Salomé Oudet

Transcription

Salomé Oudet
ANNÉE 2015
THÈSE / UNIVERSITÉ DE RENNES 1
sous le sceau de l’Université Européenne de Bretagne
pour le grade de
DOCTEUR DE L’UNIVERSITÉ DE RENNES 1
Mention : Mathématiques et applications
Ecole doctorale MATISSE
présentée par
Salomé Oudet
préparée à l’unité de recherche UMR 6625 du CNRS : IRMAR
Institut de Recherche Mathématiques de Rennes
U.F.R de Matématiques
Équations de
Thèse soutenue à Rennes
le 03/11/2015
devant le jury composé de :
Hamilton-Jacobi
Yves ACHDOU
sur des réseaux
Ariela BRIANI
ou des structures
hétérogènes
Professeur, Univ. Paris 7 / co-directeur de thèse
Maitre de conférence, Univ. François Rabelais de Tours
/ examinatrice
Nicolas FORCADEL
Professeur, INSA de Rouen / examinateur
Cyril IMBERT
Directeur de recherche CNRS, Univ. Paris-Est Créteil &
ENS Paris / rapporteur
Olivier LEY
Professeur, INSA de Rennes / examinateur
Claudio MARCHI
Maitre de conférence, Univ. de Padoue (Italie) / rapporteur
Nicoletta TCHOU
Maitre de conférence, Univ. Rennes 1 / directrice de
thèse
2
REMERCIEMENTS
En premier lieu, je tiens à remercier mes directeurs de thèse, Nicoletta Tchou et Yves Achdou, pour
m’avoir ouvert les portes de la recherche et de l’univers subtil de la théorie des solutions de viscosité.
Je leur suis très reconnaissante de m’avoir proposé ce sujet de thèse que j’ai trouvé passionnant. Leur
encadrement et leurs conseils avisés tout au long de ces trois années m’auront beaucoup appris sur les
mathématiques et la rigueur du métier d’enseignant-chercheur.
Je tiens à remercier vivement Cyril Imbert et Claudio Marchi d’avoir accepté de rapporter cette
thèse. Je les remercie pour le temps, l’énergie et le soin qu’ils auront consacrés à cette tâche. Je remercie
particulièrement Cyril Imbert avec qui j’ai eu le plaisir de discuter de mathématiques et autres sujets
à plusieurs reprises au cours de ma thèse et qui m’avait déjà fait l’honneur de parrainer les premières
Rencontres Doctorales Lebesgue en 2014. Je remercie également chaleureusement Ariela Briani, Nicolas
Forcadel et Olivier Ley qui m’ont fait le plaisir de venir compléter mon jury.
La découverte d’une spécialité des mathématiques se fait aussi par la rencontre avec les mathématiciens du domaine. De ce point de vue, mes trois années de thèse ont été bien remplies. Tout d’abord, j’ai
eu la chance de participer aux rencontres de l’ANR HJnet qui m’ont permis de côtoyer régulièrement
des éminents chercheurs de la communauté. Ces rencontres auront été une grande chance pour moi et
m’auront beaucoup apporté aussi bien sur le plan scientifique que humain. Enfin, les mathématiques
étant internationales, la thèse est aussi l’occasion de découvrir les chercheurs d’autres pays ainsi que
leur culture. Dans cette optique, j’ai eu le privilège de rencontrer Yoshikazu Giga à Tokyo, à la fin de ma
deuxième année de thèse. Je le remercie pour son accueil chaleureux, sa gentillesse et l’honneur qu’il
m’a fait en m’invitant. Je remercie également Hitoshi Ishii et Shigeaki Koike qui m’ont fait l’honneur de
m’inviter à parler dans leurs séminaires respectifs. Enfin, je remercie bien sûr Olivier Ley qui grâce à
ses qualités relationnelles, a rendu possible cette aventure.
J’ai trouvé à l’IRMAR un cadre de travail plus qu’agréable. La bonne ambiance qui règne dans ces
murs découle de la grande cohésion et convivialité de tout le personnel, tant du côté de la recherche
que du côté administratif. Je remercie d’abord, les membres des équipes d’équations aux dérivées partielles et d’analyse numérique qui m’ont réservé un chaleureux accueil dès le début de ma thèse. Plus
généralement, j’ai aimé discuter avec de nombreux chercheurs (toutes catégories confondues) de l’IRMAR à l’occasion de séminaires, colloquiums ou pauses café. Je n’oublie pas, bien sûr, les chercheurs
de l’INSA et de l’ENS qui, bien que dans d’autres murs, participent pleinement à la vie de l’IRMAR et
de ses équipes. Un grand merci aussi à la bande des doctorants et post-doctorants qui a grandement
contribué à rendre ces trois années de thèse plus qu’agréable et sans qui l’IRMAR serait bien vide. Enfin, je remercie vivement le personnel administratif et technique de l’IRMAR qui s’est montré toujours
disponible et d’une efficacité redoutable pour venir à bout de toutes mes péripéties.
L’enseignement a naturellement représenté une part très enrichissante de mon travail de doctorante.
Je remercie beaucoup Arnaud Debussche pour m’avoir donné l’opportunité d’effectuer mes missions
d’enseignement à l’ENS Rennes. Je remercie les professeurs des cours dont j’ai assuré les TD, Philippe
Chartier, Benoît Cadre et Thibaut Deheuvels, avec qui j’ai apprécié de travailler en équipe. Un grand
merci également aux élèves qui se sont avérés fort sympathiques et qui ont rendu cet enseignement très
agréable.
La thèse étant l’aboutissement d’un long cheminement, je tiens également à remercier tous les professeurs de mathématiques et autres matières qui ont cru en moi et qui m’ont encouragée dans cette
voie : mes enseignants de l’ENS, en particulier le distingué Michel Pierre, le percutant Arnaud Debussche
et le fin pédagogue Grégory Vial ; mes enseignants de l’Université Paris 13, tels le rigoureux Olivier Brinon1 , l’accueillant Alain Rousseau et surtout mon mentor, le très cher Alain Grigis ; mes enseignants
1
et sa tarte à la crème.
3
4
de classe préparatoire dont le pince-sans-rire Jean-Pierre Vial et l’engagée Marie-Christine Paris du lycée Buffon et la passionnante Nathalie Cleirec et son acolyte physicienne, Julie Fournier, du lycée Paul
Éluard de Saint-Denis.
Mon engouement pour les mathématiques et la recherche me viennent également de mes proches
et amis scientifiques que je côtoie pour certains depuis mon plus jeune âge : les passionnés Jean-Marc
et Nathalie, la créative Mireille, mon modèle, le Père Lionel ... et naturellement mon ingénieux Papi.
Ces longues années d’étude ont été jalonnées de nombreuses rencontres et je tiens vivement à remercier tous ces compagnons de route, en particulier les collègues doctorants de l’IRMAR dont Zineb,
mon excellente co-bureau et la jeune Coralie qui lui a succédé, Renan, Pierre-Yves2 , Damien et Jean-ϕ ,
mes partenaires de coinche préférés, Ophélie, Arnaud, Blandine, Julie, Florian, Gwezheneg, Alexandre,
Benois et les deux Hélène, tous partisans du tarot à 15, Charles, le militant chevronné, Axel, Néstor,
VMK, YZ, Christian, Sachio et Türkü, les fidèles du marché des Lices, Vinh, Viet et Tuyen, les camarades
HJnet, Ana Maria, Thiago, Maycol et Andres, les amis d’Amérique latine, Momo le furtif3 , Federico le
cuistot, Gabriel et Alexandre les rugbymen, Basile l’esthète, Mouton, Kodjo, Jeremy, Laurent Pater et
Thibaut, les anciens de la bande, Adrien, Marine, Grégory, Maria, Khang, Clément et Valentin, la relève,
et enfin Maxime et Mac les Ker-Lannais.
N’oublions pas la formidable promotion 2008 de l’ENS Rennes, la clique de Paris 13, Julien, Amandine et Grégory et les autres amis rencontrés ci et là, Siargey, Daria, Alexandre et Valentine parmi tant
d’autres.
Je voudrais également dire un grand merci à mes indéfectibles compères qui, contrairement à moi,
ont suivi leurs rêves d’artiste : Elise la violoniste, Sara la comédienne, Simon le funambule, Agathe,
Nico et Fanny, libres artisans de la construction insolite, et surtout Camille, la scénographe, mon alter
ego de toujours.
Je remercie évidemment mes parents pour leur soutien inconditionnel ; ils le méritent. Un petit clin
d’œil à mes deux frangines, Léa, "ma petite crotte" et Barbara, ma jumelle préférée, ainsi qu’à sa famille
de cœur, Charles, Nicolas, Mathilde et Dario, et enfin à Florianne, ma sister gabonaise. Je remercie aussi
mes grands-parents qui ont toujours suivi mon parcours avec beaucoup d’intérêt et d’affection.
Je n’oublie pas bien sûr Ninja et le Yeuv’ que je remercie sincèrement pour leur sens de l’hospitalité
et leur gentillesse.
Un merci tout particulier aux déménageurs bretons et à tous ceux chez qui j’ai squatté ces deux
derniers mois.
Enfin, un grand merci à toi, Tristan. À défaut de m’avoir convertie aux p-adiques, tu m’auras rendue
Tristan-addict. Le bon déroulement de cette thèse te doit beaucoup. Merci pour ton calme, ta bonne
humeur, ton humour, ton originalité, ta barbe rousse4 ...
2
Alias Petit-Bois.
Merci pour ton bureau qui aura si souvent été notre terrain de jeux
4
Si, si ! Démonstration : Tristan est isomorphe. Les pandas roux sont isomorphes. Donc Tristan et les pandas roux sont isomorphes. Or il est bien connu que les pandas roux sont trop mignons et ont la barbe rousse.
3
TABLE
I
DES MATIÈRES
Introduction
7
1.1 Contexte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Équations de Hamilton-Jacobi discontinues . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Problématique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 Résultats existants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Équations de Hamilton-Jacobi sur des réseaux . . . . . . . . . . . . . . . . . . . . .
1.3.1 Problématique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 Résultats antérieurs à la thèse . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3 Résultats obtenus par d’autres auteurs durant la préparation de la thèse .
1.3.4 Résumé des Chapitres 2 et 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Homogénéisation et perturbations singulières . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Problématique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 Résultats obtenus par d’autres auteurs durant la préparation de la thèse .
1.4.3 Résumé du Chapitre 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Organisation du manuscrit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
11
11
13
20
21
23
25
32
41
41
42
49
54
II Hamilton-Jacobi-Bellman equations on heterogeneous structures with geometric singularity
55
2 Hamilton-Jacobi equations for optimal control on junctions and networks
2.1 The junction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 The geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 The optimal control problem . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Test-functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Definition of viscosity solutions . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.4 Hamiltonians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.5 Existence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Properties of viscosity sub and supersolutions . . . . . . . . . . . . . . . . . . . . .
2.4 Comparison principle and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.2 Convergence of the Hamiltonian at the vertex as " → 0 . . . . . . . . . .
2.5.3 Convergence of the sub or super solutions as " → 0 . . . . . . . . . . . .
2.6 Extension to a more general framework with an additional cost at the junction
2.7 The case of a network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.1 The geometrical setting and the optimal control problem . . . . . . . . .
2.7.2 The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.3 Comparison principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7.4 Existence and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.8.1 Proof of Lemma 2.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
57
58
58
58
61
61
61
62
62
64
65
69
70
70
70
72
74
75
75
77
77
77
78
78
3 Hamilton-Jacobi equations for optimal control on multidimensional junctions
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 First case : full controllability near the interface . . . . . . . . . . . . . . . . . . .
3.2.1 Setting of the problem and basic assumptions . . . . . . . . . . . . . . . .
3.2.2 The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
81
83
83
86
5
6
TABLE DES MATIÈRES
3.2.3 Properties of viscosity sub and supersolutions . . . . . . . . . . . . . . . . . . . . . . 90
3.2.4 Comparison principle and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.3 Second case : normal controllability near interface . . . . . . . . . . . . . . . . . . . . . . . 96
3.3.1 The new framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.3.2 Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.3.3 Comparison principle and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.4 Extension to a more general framework with additional dynamics and cost at the interface109
3.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.5.1 Proof of the Proposition 3.2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.5.2 Proof of the Proposition 3.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.5.3 Proof of the Proposition 3.2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.5.4 Proof of Lemma 3.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.5.5 Proof of Theorem 3.3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
III
A singular perturbation problem
123
4 Asymptotic behaviour of Hamilton-Jacobi equations defined on two domains separated by
an oscillating interface
125
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.1 The geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.1.2 The optimal control problem in Ω"L ∪ ΩR" ∪ Γ" . . . . . . . . . . . . . . . . . . . . . . . 127
4.1.3 The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.1.4 Main result and organization of the paper . . . . . . . . . . . . . . . . . . . . . . . . 130
4.2 Straightening the geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.2.1 A change of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.2.2 The optimal control problem in the straightened geometry . . . . . . . . . . . . . . 131
4.2.3 The Hamilton-Jacobi equation in the straightened geometry . . . . . . . . . . . . . 132
4.3 Asymptotic behavior in Ω L and ΩR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.3.1 For the Hamilton-Jacobi equations in Ω L and ΩR , nothing changes . . . . . . . . . 134
4.3.2 An ingredient in the effective transmission condition on Γ : the Hamiltonian HΓ
inherited from the half-planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.4 A new Hamiltonian involved in the effective transmission condition . . . . . . . . . . . . . 135
4.4.1 Fast and slow variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.4.2 Ergodic constants for state-constrained problems in truncated domains . . . . . . 136
4.5 Further properties of the correctors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
4.6 A comparison principle for the effective equation . . . . . . . . . . . . . . . . . . . . . . . . 142
4.6.1 Properties of E(·, ·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.6.2 The comparison principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.7 Proof of the main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.7.1 A reduced set of test-functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.7.2 Proof of Theorem 4.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.8.1 Proof of Lemma 4.4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.8.2 Proof of Lemma 4.5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.8.3 Proof of Theorem 4.7.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
PREMIÈRE
PARTIE
Introduction
7
8
CHAPITRE 1
INTRODUCTION
1.1
Contexte
Ce paragraphe vise à donner un cadre général dans lequel s’inscrivent les résultats de ce mémoire, sans
prétendre à l’exhaustivité.
Les équations de Hamilton-Jacobi apparaissent dans des domaines très variés tels que la théorie du
contrôle, la théorie des jeux, le traitement d’image, les propagations de fronts en physique ou encore
les mathématiques financières. L’exemple le plus classique l’équation de Hamilton-Jacobi est l’équation
eikonale
| Du |= 1,
(1.1.1)
qui intervient en optique géométrique ou dans des problèmes de contrôle optimal de temps de sortie. Nous verrons plus tard que l’étude de problèmes de contrôle optimal conduit naturellement à des
équations de Hamilton-Jacobi-Bellman, une sous-famille des équations de Hamilton-Jacobi. L’équation
eikonale (1.1.1) est un exemple d’équation de Hamilton-Jacobi-Bellman.
La théorie des solutions de viscosité à été introduite par M. G. Crandall et P.-L. Lions en 1981
dans [CL81,Lio82,CL83] pour résoudre ces équations. Elle s’est imposée comme la bonne notion de solutions pour les équations de Hamilton-Jacobi dont l’étude était restée très difficile jusqu’alors, en raison
des fortes non-linéarités présentes. L’idée de base de la théorie des solutions de viscosité est d’utiliser
une classe de fonctions-test assez régulières (pour donner un sens aux dérivées partielles intervenant
dans l’équation) qui viennent toucher les sous-solutions (resp. sur-solutions) potentielles par dessus
(resp. par dessous). La théorie des solutions de viscosité pour les équations du premier ordre fait l’objet
de plusieurs monographies, par exemple celles de P-L. Lions [Lio82], de G. Barles [Bar94, Bar13] ou de
M. Bardi et I. Capuzzo-Dolcetta [BCD97]. En raison de son large spectre d’applications, de son élégance,
sa simplicité et sa pertinence, la théorie des solutions de viscosité s’est considérablement développée
depuis ces trente dernières d’années. Il a par exemple assez vite été compris que la théorie pouvait
s’étendre à la résolution d’équations elliptiques ou paraboliques non-linéaires, éventuellement dégénérées, du second ordre, par exemple les équations de Hamilton-Jacobi-Bellman provenant du contrôle
stochastique, ou des équations complètement non-linéaires. L’étude des équations du second ordre a fait
l’objet d’une considérable littérature, voir par exemple R. Jensen [Jen88] en 1988, R. Jensen, P.-L. Lions
et P. Souganidis [JLS88] en 1988, et la monographie de M. G. Crandall, H. Ishii et P.-L. Lions [CIL92].
De nombreux résultats et méthodes sont devenus classiques et s’appliquent à une large classe d’équations aux dérivées partielles. Dans ce mémoire, nous nous concentrerons sur l’étude d’équations de
Hamilton-Jacobi du premier ordre.
Comme ce travail concerne la théorie des solutions de viscosité pour des équations de HamiltonJacobi du premier ordre, nous rappelons ici certains résultats ou méthodes classiques pour ces équations, sachant que certaines d’entre eux s’adaptent au cas des équations du second ordre :
. Les principes de comparaison qui permettent de comparer les sous- et sur-solutions de viscosité
d’une même équation et qui impliquent en particulier le résultat d’unicité de solutions de viscosité,
voir les travaux fondateurs [CL81, Lio82, CL83] et [Bar94, Bar13, BCD97].
. Les résultats de stabilité sont intrinsèquement liés à la notion de solution de viscosité, voir
[CL81, Lio82, CL83, Bar94, Bar13, BCD97]. Ils permettent de passer à la limite dans des équations de Hamilton-Jacobi du premier ordre (ou du second ordre). La notion de stabilité est un
des arguments clefs de la méthode de la viscosité évanescente initialement utilisée pour résoudre
certaines équations de Hamilton-Jacobi du premier ordre.
9
10
CHAPITRE 1. INTRODUCTION
. Très tôt, on a compris que l’approximation des solutions des équations de Hamilton-Jacobi
pouvait être étudiée en utilisant les notions fondamentales que sont les principes de comparaison
et la stabilité. Dans ce sens, deux articles fondateurs sont dûs à M. Crandall et P-L. Lions [CL84]
et G. Barles et P. Souganidis [BS91].
. Les méthodes spécifiques à l’étude des équations de Hamilton-Jacobi-Bellman issues de problèmes de contrôle optimal ont été initiées par P.-L. Lions dans [Lio82]. Le cadre du contrôle optimal est particulier, puisqu’on sait a priori que la fonction valeur du problème de contrôle optimal
considéré est un bon candidat pour être solution de l’équation de Hamilton-Jacobi-Bellman obtenue grâce au principe de programmation dynamique. En résumé, l’existence de solutions pour
les équations de Hamilton-Jacobi-Bellman découle du principe de programmation dynamique,
et on dispose de principes de sous- et sur-optimalité. Une référence très complète traitant des
relations entre les problèmes de contrôle optimal et les équations de Hamilton-Jacobi-Bellman
est le livre de M. Bardi et I. Capuzzo-Dolcetta [BCD97]. Notons que la classe des équations de
Hamilton-Jacobi-Bellman est une sous-classe très importante des équations de Hamilton-Jacobi,
puisque grâce à la transformée de Legendre-Fenchel toute équation pouvant s’écrire sous la forme
λu(x) + H(x, Du(x)) = 0
ou
∂ t u(t, x) + H(x, Du(t, x)) = 0
u(0, x) = u0 (x)
x ∈ Ω,
dans (0, T ) × Ω,
dans Ω,
avec des Hamiltoniens H convexes par rapport à la dernière variable peut être associée à un problème de contrôle optimal.
Les travaux contenus dans ce mémoire sont liés à la question des problèmes de contrôle optimal avec des contraintes sur l’état. L’équation de Hamilton-Jacobi-Bellman associée est alors
complétée par des conditions au bord particulières d’abord introduites et étudiées par M. Soner [Son86a, Son86b] et par I. Capuzzo-Dolcetta et P-L. Lions [CDL90].
. La notion de solution de viscosité discontinue introduite par H. Ishii dans [Ish87a], voir aussi
[Ish85, Ish89] a en particulier permis l’étude d’équations de Hamilton-Jacobi du premier ordre
(ou elliptiques du second ordre ou paraboliques) avec des conditions aux limites générales. En
particulier on est capable de prouver des résultats d’unicité et de stabilité dans ce nouveau cadre.
. La méthode de Perron : H. Ishii en 1987 dans [Ish87a], a eu l’idée de généraliser la méthode
de Perron développée pour montrer l’existence de solutions classiques de l’équations de Poisson
au cadre des solutions de viscosité, pour une classe très générale d’équations de Hamilton-Jacobi
du premier ordre (ou du second ordre). On utilise généralement la méthode de Perron dans le
cadre des solutions discontinues au sens de Ishii, puis on utilise éventuellement un principe de
comparaison fort pour prouver que la solution obtenue est continue.
. L’homogénéisation, les problèmes de perturbations singulières et les problèmes de comportement en temps longs associés aux équations de Hamilton-Jacobi font l’objet d’une vaste littérature depuis le travail fondateur de P-L. Lions, G. Papanicolaou et S. Varadhan [LPV]. La méthode
de la fonction-test perturbée, introduite par L.C. Evans dans [Eva89], permet de prouver des
résultats d’homogénéisation périodique pour des équations de Hamilton-Jacobi.
La littérature sur les solutions de viscosité est maintenant énorme. Il reste pourtant des questions
ouvertes. Par exemple, comme les définitions de sous- et sur-solutions de viscosité passent par l’intermédiaire de fonctions-test régulières, la notion de solution de viscosité n’est pas totalement claire
lorsque l’équation est posée dans un domaine non ouvert et présentant des singularités géométriques.
Une autre question intéressante est de trouver une bonne notion de solution de viscosité pour des équations de Hamilton-Jacobi pour lesquelles le Hamiltonien est discontinu par rapport à la variable d’état.
En particulier, la question de donner un sens à un problème faisant intervenir différentes équations de
Hamilton-Jacobi dans différentes régions de l’espace, avec des conditions de transmission apropriées
n’est pas simple et fait l’objet d’études récentes.
Il est important de remarquer que les deux questions évoquées ci-dessus sont fortement liées. Par
exemple, si l’on considère des équations de Hamilton-Jacobi posées sur des réseaux (des sommets reliés
1.2. ÉQUATIONS DE HAMILTON-JACOBI DISCONTINUES
11
entre eux par des arêtes, penser à un graphe), il est naturel d’envisager des Hamiltoniens différents sur
chaque arête. On peut par exemple penser à des applications en trafic routier où la dynamique change
d’une route à l’autre et où les limitations du code de la route diffèrent d’une route à l’autre.
Dans les deux problèmes évoqués ci-dessus, la première question est de trouver un cadre théorique qui
permette comme pour les problèmes plus classiques d’assurer l’existence, l’unicité et la stabilité. Une
fois ce cadre bien défini, de nombreuse questions se posent naturellement, telles que l’étude de problèmes d’homogénéisation ou de comportement en temps long.
L’intérêt pour ce genre de questions a crû depuis une dizaine d’années, comme on l’a vu au congrès
international Netco de Tours en 2014. Ce mémoire s’inscrit dans cette problématique.
1.2
Équations de Hamilton-Jacobi discontinues
Dans cette partie, on essaye de donner un aperçu de l’état de l’art de la théorie des solutions de viscosité
pour des équations de Hamilton-Jacobi du premier ordre faisant intervenir des Hamiltoniens discontinus
par rapport à la variable d’état.
1.2.1
Problématique
Comme nous l’avons signalé ci-dessus, les équations de Hamilton-Jacobi interviennent dans des domaines très variés tels que la physique, la théorie du contrôle ou l’économie. Il est assez naturel d’envisager des équations de Hamilton-Jacobi pour lesquels le Hamiltonien serait discontinu par rapport à
la variable d’état. Par exemple, si l’on considère un problème de contrôle optimal posé dans un milieu
comprenant deux phases séparées par une interface, l’intuition veut que la fonction valeur de ce problème (pour peu qu’il soit bien posé) soit solution de viscosité (en un sens à définir) d’une équation
de Hamilton-Jacobi-Bellman faisant intervenir deux Hamiltoniens indépendants de part et d’autre de
la frontière, avec une condition de transmission à l’interface dont la définition dépend des limitations
qu’on impose sur les dynamiques en un point donné de l’interface. Ainsi, l’équation de Hamilton-JacobiBellman satisfaite par la fonction valeur devrait a priori être discontinue. Un autre exemple issu de
la cristallographie fait intervenir un Hamiltonien discontinu par rapport à la variable d’espace. Cette
exemple est étudié par Y. Giga et N. Hamamuki dans [GH13]. On s’intéresse à un modèle de croissance
de cristaux (décrit précisément dans [BCF51] et [SZN+ 10]) qui fait intervenir deux mécanismes physiques qui interagissent : un mécanisme de nucléation et un mécanisme de croissance en spirale. La
combinaison de ces deux mécanismes physiques ramène l’étude de la croissance du cristal dans l’espace à trois dimensions à celle d’un phénomène de nucléation bidimensionnel. Plus précisément, on est
ramené à étudier la croissance d’un cristal bidimensionnel inexistant au temps t = 0 et évoluant sur un
support plan par apport extérieur de molécules. L’apport extérieur de molécules est localisé, presque
ponctuel, et à vitesse constante. Le phénomène de nucléation bidimensionnel que l’on vient de décrire
est modélisé par une équation de Hamilton-Jacobi avec un terme source discontinu.
Les équations de Hamilton-Jacobi discontinues présentent des difficultés nouvelle du point de vue
de la théorie des solutions de viscosité. La question de l’existence n’est pas forcément un obstacle majeur
puisque la méthode de Perron s’applique en général. De même, comme les arguments classiques pour
prouver des résultats de stabilité fonctionnent pour des problèmes assez généraux, la question de la
stabilité est en général surmontable. La difficulté essentielle réside dans la preuve d’un résultat d’unicité
ou plus précisément d’un principe de comparaison. En effet, la voie classique pour prouver un principe
de comparaison est la méthode de dédoublement des variables, voir [CL81, Lio82, CL83, Bar94, Bar13,
BCD97]. Il est bien connu que la méthode standard de dédoublement des variables ne fonctionne pas
dès que le Hamiltonien est discontinu par rapport à la variable d’état. Voici un exemple très simple d’une
équation pour laquelle on ne peut pas espérer appliquer directement la méthode de dédoublement des
variables :
λu(x) + H(x, Du(x)) = 0, x ∈ R,
(1.2.1)
où λ > 0 est une constante strictement positive et où le Hamiltonien H : R × R → R est défini par
H(x, p) =
§
H L (p)
HR (p)
si x < 0,
si x ≥ 0,
12
CHAPITRE 1. INTRODUCTION
les Hamiltoniens H L et HR étant supposés convexes et coercifs. Ici le Hamiltonien H est discontinu
en x = 0. Une idée naturelle, exploitée par Barles et al [BBC13, BBC14], est d’employer la notion de
solution de viscosité au sens de Ishii pour les Hamiltoniens discontinus [Ish85]. Plus précisément, une
fonction semi-continue supérieurement (scs) u : R → R est une sous-solution de (1.2.1) si pour tout
x ∈ R et ϕ ∈ C 1 (R) tels que u − ϕ admet un maximum local en x,
λu(x) + H L (ϕ 0 (x)) ≤ 0,
λu(x) + HR (ϕ 0 (x)) ≤ 0,
λu(x) + min (H L (ϕ 0 (x)), HR (ϕ 0 (x))) ≤ 0,
si x < 0,
si x > 0,
si x = 0.
De même, une fonction semi-continue inférieurement (sci) u : R → R est une sur-solution de (1.2.1) si
pour tout x ∈ R et ϕ ∈ C 1 (R) tels que u − ϕ admet un minimum local en x,
λu(x) + H L (ϕ 0 (x)) ≥ 0,
λu(x) + HR (ϕ 0 (x)) ≥ 0,
λu(x) + max (H L (ϕ 0 (x)), HR (ϕ 0 (x))) ≥ 0,
si x < 0,
si x > 0,
si x = 0.
La méthode de dédoublement des variables classiquement utilisée pour prouver des principes de comparaison entre sous- et sur-solutions ne permet pas de conclure dans ce cadre. En effet, essayons de
montrer par la méthode de dédoublement des variables que u − v ≤ 0 sur R, si u est une sous-solution
bornée et v une sur-solution bornée. L’idée de base est d’introduire la fonction auxiliaire
φ" (x, y) = u(x) − v( y) −
|x − y|2
.
"2
(1.2.2)
On ajoute généralement d’autres termes à la fonction φ" pour assurer l’existence d’un maximum ou
pour localiser ce maximum, mais nous les oublierons pour simplifier l’exposé. On suppose donc que la
fonction φ" définie par (1.2.2) admet un maximum en un point (x " , y" ) ∈ R2 . On montre alors par des
arguments classiques que les suites x " et y" convergent vers un même point x̄ quand " → 0. Si x̄ > 0
(le cas x̄ < 0 se traitant de la même manière), quitte à supposer " suffisamment petit, on obtient par
des arguments maintenant standard l’inégalité
λ(u(x " ) − v( y" )) + HR


‹
‹
2(x " − y" )
2(x " − y" )
− HR
≤ 0.
"2
"2
Dans
de régularité du Hamiltonien HR , on peut montrer que
€ ce cas,Šd’après
€ les propriétés
Š
2(x " − y" )
2(x " − y" )
HR
− HR
→ 0 quand " → 0 et un raisonnement par l’absurde permet de conclure
"2
"2
que u − v ≥ 0 sur R.
Si au contraire x̄ = 0, alors il se peut que les suites x " et y" convergent vers 0 et que x " < 0 < y" pour
tout ". Dans ce cas, on obtient l’inégalité
λ(u(x " ) − v( y" )) + H L

‹

‹
2(x " − y" )
2(x " − y" )
− HR
≤ 0.
"2
"2
Comme les Hamiltoniens€ HR et HŠL sont€a priori Štotalement indépendants l’un de l’autre, on n’a aucun
2(x " − y" )
2(x " − y" )
espoir d’obtenir que H L
− HR
→ 0 quand " → 0 et donc de pouvoir conclure par
"2
"2
l’absurde.
Il est facile de comprendre que le problème n’est pas spécifique à cet exemple, mais apparaît quand le
Hamiltonien est discontinu par rapport à la variable d’état.
On verra aussi plus tard que dire qu’une fonction est sous- et sur-solution au sens de Ishii ne suffit pas en
général à caractériser une solution du problème. Pour caractériser la solution, il faudra généralement
imposer des conditions supplémentaires à l’interface, qu’on pourra appeler des conditions de transmission. Associées à ces conditions de transmission, il faudra proposer de nouvelles notions de solution de
viscosité et développer une nouvelle théorie pour obtenir éventuellement l’unicité soit en modifiant la
technique de dédoublement des variables, soit par d’autres méthodes.
Remarquons qu’on peut espérer obtenir des informations supplémentaires dans le cadre du contrôle optimal, c’est à dire pour des équations de Hamilton-Jacobi-Bellman discontinues. En effet, comme nous
l’avons déjà signalé, nous disposons en général de formules de représentation pour les solutions de
1.2. ÉQUATIONS DE HAMILTON-JACOBI DISCONTINUES
13
telles équations, qui sont les fonctions valeur de problèmes de contrôle optimal. Dans le cadre continu,
les sous-solutions (resp. sur-solutions) peuvent être caractérisées par des propriétés de sous-optimalité
(resp. sur-optimalité). Ainsi, il est naturel d’espérer que de telles propriétés restent vraies pour des
équations de Hamilton-Jacobi-Bellman discontinues, ce qui nous donnerait des informations supplémentaires à exploiter.
Développer une bonne théorie qui s’adapte au cas des équations de Hamilton-Jacobi discontinues par
rapport à la variable d’état présente donc un intérêt du point de vue des applications, mais également
un intérêt théorique certain. C’est une question difficile à laquelle, à notre connaissance, les premiers
éléments de réponse ont moins de vingt ans.
1.2.2
Résultats existants
Dans ce paragraphes, nous souhaitons donner un aperçu d’avancées importantes qui ont été faites à
propos des équations de Hamilton-Jacobi discontinues. Les techniques développées pour étudier ces
équations reposent sur des arguments venant soit de la théorie des équations aux dérivées partielles,
soit de la théorie du contrôle optimal. Nous allons distinguer les approches purement “EDP” de celles
mettant en jeu des arguments issus de la théorie du contrôle optimal.
1.2.2.1
Le cas particulier des équations de Hamilton-Jacobi-Bellman : combinaison de la théorie
des solutions de viscosité et de la théorie du contrôle optimal
Comme rappelé plus haut, dans le cas des équations de Hamilton-Jacobi-Bellman, nous disposons en
général de formules de représentation pour les solutions et de principes de sous- et sur-optimalité pour
les sous- et sur-solutions. Ce cadre semble donc a priori plus propice pour aborder la question des
Hamiltoniens discontinus. C’est effectivement ce type d’équations qui a été étudié le premier et pour
lequel on trouve le plus de résultats.
Le premier travail qu’il semble important d’évoquer au sujet des Hamiltoniens discontinus est l’article
de P. Dupuis [Dup92] publié en 1992. Dans cet article, l’auteur considère des problèmes de calcul des
variations avec intégrandes discontinus, plus précisément, des problèmes comportant plusieurs régions
dans lesquelles sont prescrits différents intégrandes. Après avoir donné un sens a de tels problèmes,
c’est-à-dire après avoir précisé la définition de l’intégrande aux points de jonction entre les différentes
régions, on obtient des problèmes de calcul des variations avec un intégrande a priori discontinu. P.
Dupuis a proposé un schéma numérique pour calculer la fonction valeur de tels problèmes. Il justifie sa
définition de l’intégrande aux points de jonction comme le seul choix permettant d’obtenir des propriétés
de stabilité forte sur la fonction valeur qui en résulte. Ces propriétés de stabilité de la fonction valeur
sont bien sûr essentielles pour espérer la convergence d’un schéma numérique, mais sont également
naturelles du point de vue de la modélisation.
Les premiers travaux portant effectivement sur l’étude d’équations de Hamilton-Jacobi-Bellman discontinues arrivent plus tard. On peut citer par exemple les travaux de P. Soravia [Sor02, Sor06] et
de P. Soravia avec M. Garavello [GS04, GS06] publiés entre les années 2002 et 2006. Par exemple
dans [GS04], les auteurs étudient des équations de Hamiton-Jacobi-Bellman résultant de problèmes de
contrôle optimal avec un ensemble de contrôle non-borné et un coût instantané discontinu par rapport
à la variable d’état. Plus précisément, le coût instantané est de la forme `(x, a) = h(x, a) + g(x), où h
est une fonction continue et g est une fonction Borel-mesurable. De plus il est facile de construire un
exemple de problème de contrôle optimal respectant les hypothèses ci-dessus pour lequel la fonction
valeur est discontinue. La fonction valeur d’un tel problème n’est pas l’unique solution (même au sens de
Ishii) de l’équation de Hamilton-Jacobi-Bellman associée. Les auteurs prouvent des principes de souset sur-optimalité pour les sous- et sur-solutions qui leur permettent d’obtenir des conditions nécessaires
et suffisantes pour avoir l’unicité de la solution. Dans les cas où il n’y a pas unicité, les principes de souset sur-optimalité permettent d’obtenir une caractérisation des solutions minimale et maximale.
Dans les articles évoqués ci-dessus, les Hamiltoniens ont une structure particulière mais sont Borelmesurables dans la variables d’état et on ne dispose pas d’informations précises sur l’ensemble des
points de discontinuité. Nous verrons que parmi les différentes études sur les équations de HamiltonJacobi discontinues, bon nombre se restreignent à des classes de problèmes pour lesquels on dispose
14
CHAPITRE 1. INTRODUCTION
d’informations sur l’ensemble des points de discontinuité du Hamiltonien. C’est le cas de l’article de
A. Bressan et Y. Hong [BH07] qui s’intéressent à des problèmes de contrôle optimal définis sur des
domaines stratifiables : on suppose que l’espace des états RN peut être décomposé en une union disjointe
et finie de sous-variétés de RN . Plus précisément, on dit que RN est un domaine stratifié si il existe p ∈ N
tel que
RN = M1 ∪ · · · ∪ M p ,
où les ensembles Mi vérifient les propriétés suivantes
. pour tout i = 1, · · · , p, Mi est une sous-variété de RN
. pour tout i 6= j, Mi ∩ M j = ;
. si Mi ∩ M j 6= ;, alors Mi ⊂ M j .
Sur la Figure 1.1, on présente un exemple de stratification de R2 .
M7
•
M3
M8
•
M2
M2
M6
•
M10
M1
M5
M4
•
M9
FIGURE 1.1 – Exemple de stratification de R2
Pour un domaine stratifiable, donné avec sa stratification, les auteurs considèrent des problèmes
de contrôle optimal pour lesquels la dynamique et le coût instantané sont lipschitziens sur chaque
sous-variété. Cela revient à considérer des problèmes de contrôle optimal dans RN , avec différentes
dynamiques et différents coûts instantanés dans chaque sous-variété de la stratification. L’équation de
Hamilton-Jacobi-Bellman associée à un tel problème de contrôle présente naturellement des discontinuités localisées aux points de jonction entre les différentes sous-variétés. Ce cadre permet un large
spectre d’applications. Par exemple, les problèmes étudiés par P. Dupuis dans [Dup92] entrent dans ce
cadre. A. Bressan et Y. Hong montrent que la fonction valeur est solution de viscosité d’une équation
de Hamilton-Jacobi-Bellman (discontinue) appropriée. De plus, sous certaines hypothèses de régularité
sur la fonction valeur, ils prouvent un principe de comparaison pour cette équation qui assure que la
fonction valeur est l’unique solution de cette dernière. Ce principe de comparaison est obtenu en passant
par des arguments issus de la théorie du contrôle qui permettent de montrer que toute sous-solution
(resp. sur-solution) est plus petite (resp. plus grande) que la fonction valeur.
Les articles de G. Barles, A. Briani et E. Chasseigne [BBC13, BBC14] ont eu une influence très importante sur les travaux contenus dans ce mémoire, en particulier sur les démonstrations de principes
de comparaison pour les problèmes que nous considérons. C’est la raison pour laquelle nous souhaitons détailler les méthodes développées dans ces articles. Dans [BBC13], les auteurs considèrent des
problèmes de contrôle optimal à horizon infini dans RN avec des dynamiques et des coûts instantanés
présentant des discontinuités à la traversée d’un hyperplan H . Les auteurs font des hypothèses standard
i
i
de régularité sur les dynamiques bi : Ω × Ai → RN (i = 1, 2) et les coûts instantanés `i : Ω × Ai → R
(i = 1, 2) dans chaque demi-espace, ainsi qu’une hypothèse de contrôlabilité forte en tout point de
l’espace RN . Pour que le problème de contrôle optimal évoqué ci-dessus soit bien défini, il faut préciser
la dynamique et le coût instantané sur l’interface. Le choix fait ici est relativement proche de celui de
P. Dupuis dans [Dup92], puisqu’il autorise des combinaisons convexes des dynamiques ou des coûts
instantanés issus de chaque côté de l’interface (avec éventuellement des restrictions qu’on détaillera
plus tard). L’hypothèse de contrôlabilité assure l’existence de trajectoires admissibles, et ces trajectoires
peuvent très bien rester sur l’interface H pendant un certain temps. Les auteurs introduisent alors
1.2. ÉQUATIONS DE HAMILTON-JACOBI DISCONTINUES
15
deux ensembles de trajectoires admissibles T x et T xr eg (voir ci-dessous) qui conduisent respectivement
à deux fonctions valeur U − et U + a priori différentes. Le premier espace de trajectoires admissibles T x
est l’ensemble construit en prenant toutes les trajectoires données par le système dynamique défini au
préalable (sans restriction sur les combinaisons convexes à l’interface), alors que l’espace de trajectoires
admissibles T xr eg est un sous-espace de T x construit en sélectionnant uniquement les trajectoires faisant
intervenir des stratégies dites régulières, c’est à dire faisant intervenir des combinaisons convexes de
type pousse-pousse : on n’autorise pas des combinaisons convexes de contrôles qui correspondent à des
dynamiques entrantes dans chacun des sous-domaines et dont la dynamique résultante est tangentielle
à l’interface. On remarque que grâce à l’hypothèse de contrôlabilité, les fonctions valeur U − et U + sont
continues et même lipschitziennes.
La théorie des solutions de viscosité discontinues développée par H. Ishii conduit à considérer le
problème suivant :
λu(x) + H1 (x, Du(x))
λu(x) + H2 (x, Du(x))
λu(x) + min (H1 (x, Du(x)), H2 (x, Du(x)))
λu(x) + max (H1 (x, Du(x)), H2 (x, Du(x)))
=
=
≤
≥
0
0
0
0
dans Ω1 ,
dans Ω2 ,
sur H ,
sur H ,
(1.2.3)
où pour i = 1, 2, H i : Ωi × RN → R est le Hamiltonien défini par
H i (x, p) = max (−bi (x, a).p − `i (x, a)) .
a∈Ai
(1.2.4)
Les inégalités et égalités qui apparaissent dans (1.2.3) sont à comprendre au sens de la viscosité. Par
des arguments classiques, on obtient que les fonctions valeur U − et U + sont toutes deux des solutions
de viscosité de (1.2.3), de sorte que l’unicité pour le problème (1.2.3) est a priori fausse. Cependant,
les auteurs prouvent en fait un résultat plus fin puisque en plus de montrer que U − et U + sont solutions
de (1.2.3), ils montrent que U − est sous-solution de l’équation tangentielle
λu(x) + H T (x, DH u(x)) ≤ 0
sur H ,
(1.2.5)
où le Hamiltonien tangentiels H T : H × RN −1 → R est construit en utilisant toutes les combinaisons
convexes des contrôles issus des deux sous-domaines, qui permettent de rester sur l’interface.
Les auteurs parviennent à prouver un principe de comparaison permettant de comparer les soussolutions de (1.2.3)-(1.2.5) et les sur-solutions de (1.2.3). En particulier, la fonction valeur U − est
l’unique solution de (1.2.3) qui satisfait la condition d’interface (1.2.5). La méthode développée pour
prouver ce principe de comparaison combine des arguments d’équations aux dérivées partielles et des
arguments de la théorie du contrôle optimal. Un des arguments importants pour prouver le principe de
comparaison est une alternative portant sur les sur-solutions (qui sera détaillée ultérieurement dans ce
manuscrit) qui permet d’affirmer que pour x̄ ∈ H , une sur-solution de (1.2.3) est soit une sur-solution
de l’équation tangentielle
λu(x) + H T (x, DH u(x)) ≥ 0 sur H ,
soit sur-optimale le long de certaines trajectoires localisées près de x̄. Cette alternative découle du fait
que les trajectoires partant de x̄ sont essentiellement de deux types : soit elles quittent l’interface H
(ce qui conduit à la propriété de sur-optimalité), soit elles restent un certain temps sur cet ensemble (ce
qui conduit à l’inégalité tangentielle). L’inéquation tangentielle est étudiée avec des outils venant de la
théorie des équations aux dérivées partielles alors que le principe de sur-optimalité est intrinsèquement
lié à l’aspect contrôle optimal du problème. Une conséquence directe du principe de comparaison est
que U − est la plus petite sur-solution (et même solution) de (1.2.3). Barles et al montrent enfin que U +
est la plus grande sous-solution (et même solution) de (1.2.3).
L’article [BBC14], qui est la suite de [BBC13], a pour objectif de généraliser les résultats. Les généralisations sont multiples. D’une part, les auteurs montrent que les méthodes développées dans [BBC13]
pour des problèmes de contrôle optimal à horizon infini s’adaptent très bien au cas des problèmes
de contrôle optimal à horizon fini. Par ailleurs, ils étendent les résultats à une classe plus générale
de domaines, permettant des hypersurface H de régularité W 2,∞ comme interfaces entre les deux
demi-espaces Ω1 et Ω2 . Enfin, l’hypothèse de contrôlabilité est considérablement affaiblie, puisque les
auteurs ne supposent plus que de la contrôlabilité normale (moins forte que dans [BBC13]) et seulement dans un voisinage de la jonction. Cette hypothèse affaiblie de contrôlabilité s’interprète en disant
16
CHAPITRE 1. INTRODUCTION
que pour des points proches de l’interface, il existe des trajectoires admissibles qui permettent de se
rapprocher ou de s’éloigner de l’interface. Ce cadre plus général engendre de nouvelles difficultés. En
particulier, puisque l’hypothèse de contrôlabilité a été affaiblie, les fonctions valeur U − et U + (leur
construction est analogue à celle décrite dans [BBC13]) ne sont a priori pas continues. Ainsi, les auteurs sont obligés de travailler avec des sous- et sur-solutions de viscosité discontinues, ce qui conduit
à des preuves relativement plus techniques. En passant par des arguments de localisation et de régularisation, les auteurs prouvent un principe de comparaison analogue à celui obtenu dans [BBC13],
qui implique en particulier que la fonction valeur U − est continue et est l’unique solution du système
d’équations qui intervient dans le principe de comparaison. Plus généralement, comme précédemment,
les auteurs prouvent que la fonction valeur U − est la plus petite sur-solution (et même solution) de
viscosité continue de l’équation équivalente à (1.2.3), alors que U + est la plus grande sous-solution (et
même solution) de viscosité continue de cette même équation. Enfin, la question de la stabilité, qui
avait été peu abordée dans [BBC13], est traitée dans [BBC14]. Plus précisément, les auteurs prouvent
des résultats de stabilité pour les problèmes de contrôles associés à U − et à U + . Les arguments utilisés
pour traiter le problème associé à U − sont des arguments presque standard de la théories des solutions
de viscosité alors que ceux utilisés pour traiter le problème associé à U + sont des arguments de type
contrôle optimal.
Les articles [BBC13, BBC14] proposent une approche permettant de traiter une classe relativement
large de problèmes de contrôle optimal avec discontinuité le long d’une interface. Plus précisément,
en utilisant la terminologie des domaines stratifiés introduite par A. Bressan et Y. Hong dans [BH07],
la théorie développée dans [BBC13, BBC14] permet de considérer des problèmes de contrôle optimal
assez généraux sur des domaines associés à une stratification impliquant uniquement des sous-variétés
de dimension N et N − 1, les interfaces étant des sous-variétés de dimension N − 1. Un exemple de
problème auquel on ne peut pas appliquer directement les résultats de [BBC13,BBC14] est un problème
de contrôle optimal dans R2 faisant intervenir des points de jonction triple, c’est à dire appartenant au
bord de trois régions, comme le point O dans la Figure 1.2 ci-dessous.
M4
Ω1
M1
M5
O
Ω2
M7
•
Ω3
M3
M2
M6
FIGURE 1.2 – Un exemple de géométrie avec des coins et sa stratification
En effet, les méthodes développées dans [BBC13, BBC14] reposent en partie sur l’introduction de
notions de stratégies régulières et singulières, qui ne peuvent pas être simplement définies dans cet
exemple.
Cependant, les résultats de [BBC13, BBC14] sur la fonction valeur U − (correspondant à toutes
les trajectoires relaxées) peuvent être adaptés au cadre général des domaines stratifiables. En effet,
dans [BC14], G. Barles et E. Chasseigne proposent une suite à [BBC13, BBC14] traitant des ensembles
stratifiés généraux. Plus exactement, le but de cet article est de généraliser les résultats obtenus par A.
Bressan et Y. Hong dans [BH07]. G. Barles et E. Chasseigne se placent dans le cadre de problèmes de
contrôle à horizon fini sur des domaines stratifiés relativement généraux, les restrictions faites sur le
choix du domaine stratifié portant essentiellement sur la régularité des sous-variétés qui apparaissent
dans la stratification. Les auteurs proposent une formulation du problème de contrôle optimal par le
biais d’inclusions différentielles. Rappelons que dans [BBC13, BBC14], la dynamique et le coût instantané aux points de jonction faisaient intervenir des combinaisons convexes des dynamiques et des
coût instantanés des deux régions voisines. La formulation proposée dans [BC14] couvre ce dernier
1.2. ÉQUATIONS DE HAMILTON-JACOBI DISCONTINUES
17
cas, mais aussi d’autre situations plus complexes faisant intervenir par exemple des dynamiques et
des coûts instantanés différents sur chaque sous-variété de la stratification, en particulier sur les sousvariétés de dimension k < N qui constituent l’interface. Enfin, les auteurs font des hypothèses de
contrôlabilité normale similaires à celles proposées dans [BBC14]. Dans ce contexte, G. Barles et E.
Chasseigne introduisent une fonction valeur U qui peut être vue comme l’analogue de la fonction valeur U − de [BBC13,BBC14]. Comme dans [BBC13,BBC14], les auteurs montrent que la fonction valeur
U est solution d’une équation de Hamilton-Jacobi-Bellman faisant intervenir des inégalités de viscosité
relativement standard et des inégalités de type sous-solutions moins standard impliquant des Hamiltoniens tangentiels. Les auteurs prouvent un principe de comparaison général, de sorte que la fonction
valeur U est bien caractérisée. Une des améliorations notables par rapport à [BH07] est qu’il n’est pas
nécessaire de supposer que la fonction valeur soit 1/2-Hölderienne et que sa restriction à chaque sousvariété soit différentiable presque partout pour obtenir le principe de comparaison. De plus, le principe
de comparaison est valable pour des sous- et sur-solutions semi-continues supérieurement et inférieurement, alors que dans [BH07] il s’appliquait à des sous-solutions lipschitziennes et des sur-solutions
continues. La preuve du principe de comparaison proposée dans [BC14] se fait par récurrence sur la dimension des sous-variétés qui apparaissent dans la stratification. L’argument clef dans cette récurrence
est un résultat appelé le lemme magique, qui est analogue à l’alternative portant sur les sur-solutions
dans [BBC13, BBC14] et évoquée plus haut. Enfin, les auteurs prouvent un résultat de stabilité permettant de faire dépendre la stratification d’un paramètre ". La question de la stabilité n’était pas traitée
dans [BH07].
Les articles récents de Z. Rao et H. Zidani [RZ13] et de Z. Rao, A. Siconolfi et H. Zidani [RSZ14]
s’inscrivent également dans la lignée de l’article précurseur de A. Bressan et Y. Hong [BH07].
Dans [RZ13], Z. Rao et H. Zidani se proposent d’étudier des équations de Hamilton-Jacobi-Bellman
associées à des problèmes de contrôle optimal à horizon fini et définies sur des multi-domaines assez
généraux. Par multi-domaine, on entend domaine stratifié pour lequel on distingue des sous-régions Ωi
(les sous-variétés de dimension N de la stratification) et les morceaux de frontières Γi (les sous-variétés
de dimension strictement inférieure à N de la stratification). Les problèmes étudiés sont proches de
ceux considérés dans [Dup92] et [BBC13, BBC14]. En effet, les auteurs s’intéressent à des problèmes
où l’on considère différentes équations de Hamilton-Jacobi-Bellman dans chaque région Ωi de l’espace,
sans préciser les équations aux points de jonction. Plus précisément, les équations prescrites sur les
ensembles Ωi sont de la forme
−∂ t u(t, x) + H i (x, Du(t, x)) = 0
u(T, x) = ϕ(x)
dans (0, T ) × Ωi ,
dans Ωi ,
(1.2.6)
où ϕ : RN → R est une fonction donnée et le Hamiltonien H i : Ωi × RN → R est défini par
H i (x, p) = max (−q.p) ,
(1.2.7)
q∈Fi (x)
où Fi : Ωi → P (RN ) est une multi-fonction supposée lipschitzienne (par rapport à la métrique de
Hausdorff), à valeurs dans l’ensemble des parties compactes et convexes de RN . On peut remarquer
que pour pouvoir traiter une classe aussi générale de multi-domaines, les auteurs se restreignent à
des Hamiltoniens H i , définis par (1.2.7), correspondant à des problèmes de contrôle optimal faisant
intervenir une dynamique donnée par la multi-fonction Fi et un coût instantané nul (problème de
Mayer). Z. Rao et H. Zidani souhaitent déterminer de bonnes conditions de jonction sur les ensembles
Γi à ajouter à (1.2.6) pour obtenir l’existence et l’unicité d’une solution. L’approche proposée ici est
différente de celle proposée par G. Barles et al dans [BBC13, BBC14] car les auteurs n’imposent pas a
priori de condition de type Ishii (voir les deux dernières lignes de (1.2.3)) aux points de jonction. Une
première idée pour obtenir une condition d’interface satisfaisante est de suivre l’approche proposée
dans [BBC13, BBC14] en faisant intervenir des Hamiltoniens tangentiels HΓT , sur chaque sous-variété
j
Γ j qui constitue l’interface, le Hamiltonien HΓT étant construit à partir des combinaisons convexes de
j
dynamiques issues des régions Ωi adjacentes qui permettent de rester dans Γ j . En d’autres termes, si
pour x ∈ Γ j on note I(x) = {i : x ∈ Ωi } l’ensemble des indices correspondants à des régions adjacentes
à x et TΓ j (x) l’espace tangent à Γ j en x, le Hamiltonien HΓT est construit à partir des dynamiques
j
18
CHAPITRE 1. INTRODUCTION
e j (x) = co ∪i∈I(x) Fi (x) ∩ TΓ (x), c’est-à-dire
G
j
HΓT (x, p) = max (−q.p) .
j
(1.2.8)
e j (x)
q∈G
Les auteurs proposent une condition de jonction faisant intervenir un Hamiltonien H E , dit Hamiltonien essentiel, construit à partir d’un ensemble de dynamiques essentielles F E (x) qui contient l’ene j (x) évoqué ci-dessus, mais également les dynamiques issues de chaque région adjacente Ωi
semble G
qui pointent vers Ωi . Le concept de Hamiltonien essentiel a été introduit pour la première fois par R.C.
Barnard et P.R. Wolenski dans [BW13]. Finalement, les auteurs considèrent le problème suivant
−∂ t u(t, x) + H i (x, Du(t, x)) = 0
−∂ t u(t, x) + H E (x, Du(t, x)) ≥ 0
−∂ t u(t, x) + H E (x, Du(t, x)) ≤ 0
u(T, x) = ϕ(x)
dans (0, T ) × Ωi ,
dans (0, T ) × Γ j ,
dans (0, T ) × Γ j ,
dans Ωi ,
(1.2.9)
où les deuxièmes et troisièmes lignes contiennent les conditions d’interface. Nous avons délibérément
écrit distinctement les inégalités de sous- et sur-solution sur Γ j afin d’insister sur le fait que les définitions
de sous- et sur-solution proposées par les auteurs ici sont très différentes. En effet, l’inégalité de sursolution met en jeu des fonctions-test φ ∈ C 1 (RN ), alors que l’inégalité de sous-solution utilise des
fonctions-test φ ∈ C 1 (Γ j ) (telles que (u − φ)|Γ j admet un maximum local en x ∈ Γ j ). Les auteurs
prouvent que si la fonction ϕ est lipschitzienne, alors l’équation (1.2.9) admet une unique solution
de viscosité qui est la fonction valeur d’un certain problème de contrôle optimal. L’unicité découle
d’un principe de comparaison, valable pour des sous-solutions et sur-solutions continues. La preuve du
principe de comparaison repose sur des arguments de contrôle optimal, i.e. des propriétés de sous- et
sur-optimalité des sous- et sur-solutions.
Dans [RSZ14], les auteurs se proposent d’étudier des problèmes de contrôle optimal à horizon infini
sur des multi-domaines. Dans cette article, les auteurs se restreignent à considérer des multi-domaines
de la forme
R N = Ω 1 ∪ Ω2 ∪ Γ ,
Ω1 et Ω2 étant des sous-variétés (non-nécessairement connexes) de dimension N et Γ étant une sousvariété de dimension N − 1 supposée de classe C 2 . Comme dans [RZ13], on part d’un problème bien
défini dans les régions Ωi et il s’agit de trouver de bonnes conditions sur l’interface Γ pour que le
i
problème soit bien posé. Plus précisément, on connait les dynamiques bi : Ω × Ai → RN et les coûts insi
tantanés `i : Ω × Ai → R dans chaque région Ωi . Les hypothèses faites sur les fonction bi et `i , i = 1, 2,
sont relativement standard. Comme dans [Dup92] et [BBC13,BBC14] les auteurs choisissent de définir
la dynamique et le coût instantané aux points de jonction en faisant des combinaisons convexes des
dynamiques et des coûts instantanés issus des régions adjacentes. On peut remarquer que les restrictions faites sur le choix du domaine stratifié ont permis d’affaiblir de nombreuses hypothèses faites
dans [RZ13] de manière significative. En particulier, le coût instantané du problème de contrôle optimal n’est plus supposé nul et les hypothèses de contrôlabilité ont été considérablement réduites. En
effet, dans [RZ13] les auteurs supposaient de la contrôlabilité forte dans tout l’espace (pour assurer
la lipschitziennité de la fonction valeur). Ici les hypothèses de contrôlabilité sont localisées aux points
de jonction et sont plus faibles. Sans donner de manière précise les hypothèses de contrôlabilité faites
dans [RSZ14], on peut souligner qu’elles sont très différentes de celles utilisées dans les différents articles évoqués jusqu’ici, puisqu’elles donnent principalement des informations sur le comportement de la
solution dans la direction tangentielle à l’interface. Dans [BBC14] et [BC14], les hypothèses de contrôlabilité portaient plutôt sur la direction normale à l’interface. L’équation proposée ici pour caractériser
la fonction valeur s’écrit
λu(x) + H i (x, Du(x)) = 0
λu(x) + max (H1 (x, Du(x)), H2 (x, Du(x))) ≥ 0
λu(x) + HΓT (x, Du(x)) ≤ 0
dans Ωi ,
dans Γ ,
dans Γ ,
(1.2.10)
où pour i = 1, 2 le Hamiltonien H i est le Hamiltonien standard défini par (1.2.4) et où le Hamiltonien
HΓT est construit en sélectionnant les combinaisons convexes de dynamiques et coûts instantanés issus
des régions adjacentes qui permettent de rester sur Γ . Dans l’équation ci-dessus (1.2.10), la condition
1.2. ÉQUATIONS DE HAMILTON-JACOBI DISCONTINUES
19
d’interface apparaît sur les deuxièmes et troisièmes lignes. Les deux premières lignes de l’équation
(1.2.10) sont à comprendre en un sens usuel. En revanche l’inégalité de sous-solution sur Γ est moins
standard puisqu’elle met en jeu des fonctions-test φ ∈ C 1 (Γ ) (telle que (u − φ)|Γ admet un maximum
local en x ∈ Γ ). Les auteurs prouvent que la fonction valeur du problème de contrôle optimal considéré
ici peut être caractérisée comme étant l’unique solution de (1.2.10). L’unicité découle d’un principe de
comparaison obtenu comme dans [RZ13] par des arguments de contrôle optimal. On peut remarquer
que le principe de comparaison obtenu ici est plus général que celui obtenu dans [RZ13] puisqu’il
s’applique à des sous-solutions continues aux points de jonction (et plus sur tout l’espace) et à des
sur-solutions semi-continues inférieurement.
1.2.2.2
Approches purement EDP
Dans cette partie, nous présentons des articles portant sur l’étude d’équations de Hamilton-Jacobi discontinues non nécessairement issues de problèmes de contrôle optimal et dans lesquels les techniques
développées relèvent uniquement de la théorie des équations des dérivées partielles.
Les articles de F. Camilli et A. Siconolfi [CS03, CS05] proposent une théorie des équations de
Hamilton-Jacobi avec des Hamiltoniens seulement mesurables par rapport à la variable d’état et convexes
(ou quasi-convexes dans le cas autonome) coercifs par rapport à la variable associée à Du.
Détaillons la démarche suivie pour obtenir la notion de solution dans [CS03]. L’équation étudiée par
les auteurs se présente sous la forme
H(x, Du(x)) = 0
dans Ω,
où Ω est un ouvert non vide de RN . Dans le cas où le Hamiltonien H est continu, on dispose de la
formule intégrale
L( y, x) = inf
¨Z
1
«
0
σ(ξ(t), ξ (t))d t : ξ ∈ A y,x
pour x, y ∈ RN ,
(1.2.11)
0
de laquelle on peut déduire toute une famille de solutions fondamentales. Dans la formule (1.2.11),
σ est une fonction associée au Hamiltonien H et A y,x désigne l’ensemble des chemins ξ : [0, 1] → Ω
lipschitziens qui relient y à x. Cette formule ne peut pas être appliquée directement dans le contexte
actuel du fait que les chemins ξ ∈ A y,x sur lesquels porte l’intégration dans (1.2.11) sont de mesure
nulle par rapport à la mesure de Lebesgue. Pour remédier à ce problème, les auteurs modifient la
formule intégrale (1.2.11) en sélectionnant des sous-ensembles de A y,x . Plus précisément, la formule
intégrale proposée par les auteurs fait intervenir les ensembles
E
A y,x
= ξ ∈ A y,x : ξ et E sont transversaux ,
où E est un ensemble de mesure nulle. Un chemin ξ ∈ A y,x et un ensemble E de mesure nulle sont dit
transversaux si
µ1 ({t : ξ(t) ∈ E}) = 0,
où µ1 désigne la mesure de Lebesgue de R. Les notions de sous- et sur-solution de viscosité proposées ici sont déduites de la formulation intégrale et sont non symétriques. Remarquons que lorsque le
Hamiltonien H est continu, une solution de viscosité au sens de Crandall et Lions n’est pas forcément
une solution en le sens défini ici. Les auteurs parviennent à prouver l’existence ainsi qu’un principe de
comparaison dans ce cadre. La question de la stabilité est également abordée.
Un autre article important est celui de G. M. Coclite et N. H. Risebro [CR07] dans lequel les auteurs
étudient des équations de Hamilton-Jacobi discontinues qui peuvent apparaître dans l’étude de problèmes de reconstruction 3D à partir des ombres (connus sous le nom de "shape from shading"). Plus
précisément, G. M. Coclite et N. H. Risebro étudient des problème de Cauchy de la forme
∂ t u(t, x) + H(g(t), a(x), ∂ x u(t, x)) = 0
u(0, x) = u0 (x)
dans (0, T ) × R,
sur R,
20
CHAPITRE 1. INTRODUCTION
où le Hamiltonien H est supposé continu et les fonctions a et g sont C 1 par morceaux. De plus, les
auteurs supposent que les fonctions a et g admettent un nombre fini de point de discontinuité, de sorte
ces derniers soient localisé dans un voisinage de 0. Les auteurs proposent des notions de sous- et sursolution de viscosité proches de celles de H. Ishii, pour des fonctions uniformément continues. Dans ce
cadre, ils parviennent à prouver l’existence et l’unicité d’une solution. Pour obtenir l’unicité, les auteurs
prouvent un principe de comparaison par la méthode de dédoublement des variables, avec un choix
du terme de pénalisation qui permet d’éviter les points de discontinuité et ainsi conclure de manière
presque standard. La preuve de l’existence utilise une correspondance entre l’équation de HamiltonJacobi étudiée et une loi de conservation scalaire. En effet, pour les lois de conservation scalaire, la
méthode de front tracking permet d’approcher la solution entropique. L’intérêt de cette méthode est
qu’elle s’adapte très bien aux lois de conservation avec flux discontinus. Ainsi, en utilisant la correspondance évoquée ci-dessus et cette méthode de front tracking, les auteurs parviennent à construire un
schéma convergent pour l’équation de Hamilton-Jacobi et ainsi d’assurer l’existence d’une solution de
viscosité.
Dans la note [GGR11] Y. Giga, P. Górka et P. Rybka étudient des équations de Hamilton-Jacobi discontinues qui peuvent apparaître dans l’étude de flots par courbure moyenne singuliers. Plus précisément,
les auteurs s’intéressent à des équations de la forme
∂ t u(t, x) + H(t, x, u(t, x), ∂ x u(t, x)) = 0
u(0, x) = u0 (x)
dans (0, T ) × R,
sur R,
pour lesquelles le Hamiltonien dépend de manière non-triviale de u et est discontinu par rapport à x
le long d’une ligne Γ = {(t, r0 (t)) : t ∈ [0, T ]} ⊂ [0, T ] × R. Les auteurs imposent un cadre volontairement restrictif, avec de nombreuse hypothèses sur le Hamiltonien H, afin de pouvoir exposer une
preuve simple d’un principe de comparaison qui pourrait éventuellement être adapté à des cadres plus
généraux. Les auteurs utilisent les définitions de sous- et sur-solutions de H. Ishii. Le principe de comparaison proposé est valable pour des sous- et sur-solutions uniformément continues, avec quelques
hypothèses additionnelles sur la sur-solution afin de contrôler son comportement à l’infini. La preuve
du principe de comparaison repose essentiellement sur une perturbation de la sur-solution qui permet
d’obtenir une sur-solution stricte, et sur un argument de régularisation du Hamiltonien (rendu possible
par le fait qu’on travaille maintenant avec une sur-solution stricte) qui permet de se ramener à la théorie
classique.
Enfin, dans [GH13] Y. Giga et N. Hamamuki étudient des équations de Hamilton-Jacobi faisant
intervenir un terme source discontinu, qui apparaissent par exemple dans des modèles de formation de
cristaux. Plus précisément, les auteurs considèrent des équations de Hamilton-Jacobi de la forme
∂ t u(t, x) + H0 (x, Dx u) = r(x)
u(0, x) = u0 (x)
dans (0, T ) × RN ,
sur RN ,
où le Hamiltonien H0 : RN × RN → R est continu et coercif et la fonction r : RN → R est bornée et
semi-continue inférieurement, de sorte que le Hamiltonien H(x, p) = H0 (x, p) − r(x) est semi-continu
supérieurement par rapport à la variable x. Les auteurs proposent une notion de solution enveloppe
associée à une notion de sous-solution standard et à une notion de sur-solution enveloppe plus compliquée. Les auteurs justifient soigneusement leur définition de solution enveloppe, en discutant sur
un exemple des défauts d’existence ou d’unicité que l’on peut observer avec différentes définitions de
solution. Les auteurs parviennent à prouver l’existence et l’unicité (via un principe de comparaison)
d’une solution enveloppe pour toute condition initiale u0 uniformément continue. De plus, ils prouvent
que cette unique solution hérite de la régularité de la condition initiale u0 . La preuve du principe de
comparaison fait intervenir une méthode de dédoublement des variables (qui fonctionne ici puisque le
Hamiltonien H est semi-continu supérieurement et grâce aux choix judicieux des notions de sous- et
sur-solution qui ont été faits) et la preuve de l’existence se fait par des techniques d’approximation.
1.3
Équations de Hamilton-Jacobi sur des réseaux
L’étude des équations de Hamilton-Jabobi posées sur des réseaux ou des multi-réseaux comporte deux
types de difficultés, liées d’une part à la géométrie inhabituelle et d’autre part aux discontinuités des
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
21
données puisque dans l’exemple des réseaux, les équations diffèrent généralement d’une branche du
réseau à une autre. Ainsi, nous allons encore être confrontés à des équations de Hamilton-Jacobi discontinues, mais ces dernières sont maintenant posées sur des ensembles dont la géométrie est singulière,
et les discontinuités seront localisées aux singularités géométriques. Nous allons tenter de donner un
aperçu des travaux portant sur les équations de Hamilton-Jacobi définies sur des structures hétérogènes,
c’est-à-dire dans des ensembles comportant des singularités, et avec des Hamiltoniens discontinus par
rapport à la variable d’état.
1.3.1
Problématique
Classiquement, la notion de solution de viscosité pour des équations de Hamilton-Jacobi met en jeu
des fonctions-test assez régulières pour que toutes les dérivées partielles intervenant dans l’équation
aient un sens. Typiquement, pour une équation de Hamilton-Jacobi du premier ordre posée dans un
ouvert Ω de RN , les fonctions-test appartiendront à C 1 (Ω). De manière générale, l’étude d’équations
de Hamilton-Jacobi dans un ensemble D par des méthodes de viscosité requiert une notion de différentiabilité dans D. Par exemple, on comprend aisément comment étendre la notion classique de solution
de viscosité à des équations posées sur des sous-variétés différentiables de RN . Pour un ensemble D
seulement régulier par morceaux, la notion de différentiabilité n’est d’une part pas bien définie aux
points appartenant aux singularités géométriques, et d’autre part, les équations posées en ces points
peuvent être vues comme des conditions de transmission entre les parties régulières du domaine. Une
première étape est donc de définir une classe de fonctions-test adaptée à la géométrie de l’ensemble
D. Par exemple si D est le réseau représenté Figure 1.3, le point de jonction O entre les différentes
arêtes est une singularité géométrique, et on ne peut pas définir la notion de différentiabilité en O. En
revanche, on peut définir la notion de dérivée directionnelle d’une fonction au point O dans chacune
des directions données par les arêtes. Une classe de fonctions-test adaptée à cet exemple contiendra les
fonctions continues sur D et dont la restriction à chacune des arêtes fermées en O est de classe C 1 .
• P2
• P3
•
P4
•
O
•
P1
• PN
FIGURE 1.3 – Un réseau ne comportant qu’un point de jonction
De manière générale, sous des hypothèses naturelles sur la structure géométrique de l’ensemble
considéré (par exemples en supposant que l’ensemble est constitué d’un assemblage de sous-variétés
de dimension M et que les singularités géométriques sont concentrées sur des sous-variétés régulières
et disjointes, de dimension M − 1), on trouve aisément un espace des fonctions-test. Il reste ensuite à
proposer des conditions de transmissions et à définir une notion de solution de viscosité adaptée.
Pour comprendre les difficultés posées par les géométries singulières, il est utile de se concentrer
sur des problèmes de contrôle optimal. Nous souhaitons montrer dans un premier temps que pour un
problème de contrôle optimal posé dans une partie D de RN d’intérieur vide (c’est le cas d’un réseau),
il n’est pas souhaitable de considérer le problème comme un problème de contrôle optimal posé dans
RN avec contrainte d’état dans D.
Considérons d’abord un problème de contrôle optimal à horizon infini où l’état est contraint à appartenir
à Ω, où Ω est un ouvert non vide de RN à bord régulier. Si le facteur d’actualisation est λ, la dynamique
est donnée par f : RN × A → RN et le coût instantané par ` : RN × A → R, la fonction valeur est l’unique
solution du problème suivant
λu(x) + supa∈A (− f (x, a).Du(x) − `(x, a))
λu(x) + supa∈A (− f (x, a).Du(x) − `(x, a))
≤
≥
0 dans Ω,
0 dans Ω,
(1.3.1)
22
CHAPITRE 1. INTRODUCTION
où les inégalités sont comprises au sens de viscosité. Considérons maintenant un problème de contrôle
optimal à horizon infini posé dans une partie D de RN d’intérieur vide, comme le réseau représenté
dans la Figure 1.3. Même s’il était possible de prolonger la dynamique et la fonction coût à RN , il ne
serait pas souhaitable de voir le problème comme un problème de contrôle optimal dans RN où l’état
serait contraint à appartenir à D : en effet, la transposition de la première ligne de (1.3.1) devient alors
totalement vide d’information car l’intérieur de D est vide.
Il est donc préférable de munir D de la topologie induite par la topologie euclidienne de RN . Dans cette
topologie, D n’est pas d’intérieur vide et on peut espérer écrire des inégalités analogues à la première
ligne de (1.3.1) qui soient pertinentes (elles seront de plus particulières aux points singuliers contenus
dans l’intérieur de D).
•
•
•
•
•
•
(a)
(b)
(c)
•
(d)
(e)
FIGURE 1.4 – Exemples d’espaces ramifiés
Précisons maintenant les différents types de géométries étudiés dans cette thèse. Dans la littérature,
le terme espace ramifié est utilisé dans différents contextes et n’a pas toujours la même signification. Ici,
sauf indication contraire, on désignera par espace ramifié un sous-ensemble connexe et fermé de RN ,
obtenu comme union finie de sous-variétés disjointes de RN de dimensions strictement plus petites que
N . Quelques exemples d’espaces ramifiés sont représentés dans la Figure 1.4. Nous nous intéresserons
plus particulièrement à la sous-classe de l’ensemble des espaces ramifiés constitués d’ensembles que
nous appellerons réseaux ou multi-réseaux, le terme réseau étant plutôt réservé à des assemblages de
segments de droites ou d’arcs. Un réseau est un graphe plongé dans un espace euclidien de dimension
N , constitué de sommets (sous-variétés de dimension 0) et d’arêtes (sous-variétés de dimension 1).
Nous appellerons multi-réseau un espace ramifié dont la décomposition fait intervenir uniquement
des sous-variétés de dimension p et p − 1 pour un entier p ∈ {1, · · · , N − 1}. Un réseau est donc un
multi-réseau avec p = 1. Pour un multi-réseau quelconque, les sous-variétés de dimension p − 1 (qui
correspondent aux sommets des réseaux) seront appelées des jonctions ou interfaces, alors que les sousvariétés de dimension p (qui correspondent aux arêtes des réseaux) seront appelées des régions. Dans
la Figure 1.4, la sous-figure (a) représente un réseau et les sous-figures (a), (b) et (c) représentent des
exemples de multi-réseaux. Les espaces ramifiés représentés dans les sous-figures (d) et (e) ne sont
pas des multi-réseaux. La définition d’espace ramifié n’est pas sans rappeler la définition de domaine
stratifié introduite par A. Bressan et Y. Hong dans [BH07], également utilisée dans les articles [BBC13,
BBC14,BC14,RZ13,RSZ14] évoqués dans le paragraphe § 1.2.2. Rappelons qu’on parle de stratification
de RN , lorsque RN peut être décomposé en une union finie de sous-variétés de RN . Ainsi, les espaces
ramifiés sont construits comme les domaines stratifiés, mais on interdit les sous-variétés de dimension
N . Une conséquence immédiate de cette construction est qu’un espace ramifié est d’intérieur vide dans
la topologie de RN et présente en général des singularités géométriques.
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
23
Dans un multi-réseau ou dans un domaine ramifié, il est naturel de considérer des équations de
Hamilton-Jacobi où les Hamiltoniens sont définis de manière indépendante dans les différentes régions. En particulier, pour des problèmes de contrôle optimal posés dans un multi-réseau, il faudra
proposer des équations ou inéquations à vérifier aux interfaces. Dans tous les cas, on retrouve la difficulté, déjà vue dans des problèmes avec Hamiltoniens discontinus par rapport à la variable d’état, de
trouver une bonne notion de solution de viscosité aux interfaces qui permette en particulier d’assurer
l’existence et l’unicité de solutions et la stabilité. On rappelle qu’on ne pourra pas appliquer la méthode
de dédoublement des variables standard pour obtenir un principe de comparaison.
L’étude d’équations de Hamilton-Jacobi sur des réseaux présente en particulier un intérêt du point
du vue des applications en trafic routier. En effet, il existe des modèles de trafic routier faisant intervenir
des équations de Hamilton-Jacobi du premier ordre sur des réseaux. Du point de vue du trafic routier,
le fait d’autoriser différentes équations sur chaque arrête permettrait de prendre en compte le fait que
les dynamiques et les règles de circulation peuvent changer d’une route à l’autre.
1.3.2
Résultats antérieurs à la thèse
Il existe une vaste littérature portant sur l’étude d’équations de Hamilton-Jacobi issues de problèmes
de contrôle optimal où l’état est contraint à appartenir à la fermeture d’un ouvert de RN , voir [Son86a,
Son86b,CDL90,IK96,BCD97]. Les équations de Hamilton-Jacobi dans des ensembles fermés d’intérieur
vide n’ont été étudiées que plus récemment : en particulier, H. Frankowska et S. Plaskacz proposent
dans [FP00b, FP00a] des résultats pouvant s’appliquer à certains ensembles fermés d’intérieur vide,
mais pas à des espaces ramifiés généraux.
1.3.2.1
Premiers travaux sur les réseaux
Les réseaux constituent le plus simple exemple d’espace ramifié que l’on puisse imaginer. Il est donc
naturel que les premiers travaux concernant l’étude d’équations de Hamilton-Jacobi sur des espaces
ramifiés se sont portés sur les réseaux. Nous allons évoquer ici les trois articles sur les réseaux, [SC13,
ACCT13, IMZ13], écrits simultanément et publiés en 2013.
Dans [SC13] D. Schieborn et F. Camilli étudient des équations eikonales sur des réseaux. Plus précisément, les auteurs s’intéressent à des équations de la forme
H(x, Du(x)) = 0
dans G
(1.3.2)
où G est un réseau, H est une application de G¯ × R dans R. Si x n’est pas un sommet, Du(x) représente la dérivée directionnelle le long de l’arête contenant x préalablement orientée. Le Hamiltonien H est supposé continu par rapport à la variable d’état, ce qui veut dire en particulier que si
O est un sommet et Ji , J j deux arêtes adjacentes à O orientées de la manière suivante : l’une des
deux arêtes orientée pointe vers O tandis que l’autre ne pointe pas vers O, alors pour tout réel p,
lim x∈Ji ,x→O H(x, p) = lim x∈J j ,x→O H(x, p). Cette hypothèse supprime la principale difficulté mentionnée ci-dessus et permet aux auteurs de choisir les fonctions-test comme étant continues avec des restrictions aux arêtes fermées de classe C 1 , et satisfaisant de plus une condition de compatibilité aux
points de jonction portant sur les dérivées directionnelles. Les fonctions-test n’ont donc en fait qu’une
seule dérivée aux sommets quitte à changer l’orientation des arêtes et ressemblent donc fortement à
des fonctions C 1 . Les auteurs proposent alors une notion de solution de viscosité de (1.3.2) mettant
en jeu cette classe restreinte de fonctions-test, et parviennent à prouver l’existence, l’unicité ainsi qu’un
résultat de stabilité. La preuve de l’existence repose sur une formule de représentation impliquant la
distance géodésique. L’unicité découle d’un principe de comparaison obtenu en adaptant la méthode
spécifique aux équations eikonales développée par H. Ishii dans [Ish87b]. Les méthodes développées
dans [SC13] sont donc très spécifiques au cas de l’équation eikonale et ne peuvent pas être utilisées
pour des équations plus générales. En particulier, la classe des fonctions-test est trop petite pour traiter
une équation impliquant des Hamiltoniens discontinus aux sommets du réseaux.
Dans [ACCT13] Y. Achdou, F. Camilli, A. Cutrì et N. Tchou s’intéressent à l’étude de problèmes de
contrôle optimal avec horizon infini, pour lesquels les trajectoires sont contraintes à rester dans un
24
CHAPITRE 1. INTRODUCTION
réseau G vu comme un sous-ensemble de R2 . La dynamique est f : R2 × A → R2 et le coût instantané
est ` : R2 × A → R. Les auteurs font une hypothèse de contrôlabilité forte et supposent que l’ensemble
des contrôles admissibles A x en un point x ∈ G est un compact non vide, et que cet ensemble est
invariant le long d’une arête ouverte. Même si les dynamiques et les coûts sont continus en fonction de
x ∈ R2 , le problème conduit à une équation de Hamilton-Jacobi-Bellman discontinue par rapport à la
variable d’état, les discontinuités ayant lieu aux points de jonction du réseaux. En effet, si x n’est pas
un point de jonction du réseau et appartient à Ji , la fonction valeur satisfait une équation de HamiltonJacobi-Bellman dont le Hamiltonien est
H i (x, p) = max (− f (x, a).p − `(x, a)) ,
a∈A x
où A x est l’ensemble des contrôles admissibles au point x et p est réel représentant la dérivée directionnelle le long de l’arête contenant x. Comme le réseau G présente une singularité en tout point de
jonction, l’ensemble des contrôles admissibles A x est discontinu en chacun de ces points.
Prenons pour simplifier un réseau ne comportant qu’un point de jonction O et N arêtes Ji comme dans le
Figure 1.3. Les auteurs proposent une équation en O dont le Hamiltonien met en jeu toutes les dérivées
directionnelles dans les directions des N arêtes issues de O, puisque l’ensemble des contrôles admissibles en O est l’union des ensembles des contrôles permettant aux trajectoires issues de O d’entrer
dans l’arête Ji , pour i = 1, . . . , N . Il est remarquable que le Hamiltonien au point O puisse s’exprimer
en fonction des parties décroissantes des Hamiltoniens H i associés aux branches Ji , que Y. Achdou et
al. notent H i+ , le signe + s’expliquant par le fait que seules les dynamiques entrantes dans Ji à partir de
O sont prises en compte.
Les auteurs choisissent les fonctions-test comme étant continues sur G avec des restrictions aux arêtes
fermées J̄i de classe C 1 (J̄i ). Ils peuvent alors définir la notion de solutions de viscosité de l’équation
mentionnée ci-dessus. Remarquons que si le réseau est constitué de deux demi-droites parallèle dont
l’union est la droite réelle, la notion de solution de viscosité proposée est équivalente à la notion usuelle.
En utilisant le principe de programmation dynamique, les auteurs prouvent que la fonction valeur est
une solution de viscosité de cette équation. Pour l’unicité, les auteurs font l’hypothèse supplémentaire
que les minima des Hamiltoniens associés à chaque arête prennent la même valeur au point O. Nous
verrons dans le premier travail présenté dans ce mémoire que cette hypothèse n’est en fait pas nécessaire. Sous ces hypothèses, les auteurs proposent une preuve du principe de comparaison qui repose sur
une adaptation de la méthode de dédoublement des variables, en remplaçant le terme de pénalisation
|x− y|2
d̃(x, y)2
usuel " par " , où d̃(·, ·) est une généralisation non-symétrique de la distance géodésique. Plus
précisément, pour x, y ∈ G , d̃(x, y) représente le temps minimal pour aller du point x au point y, en
restant dans G , avec l’ensemble des dynamiques f (O, A). Les auteurs prouvent le principe de comparaison d’abord dans un cas simple où le coût instantané est indépendant du contrôle (`(x, a) = `(x)),
puis dans un cadre plus général avec des hypothèses supplémentaires.
Dans [IMZ13], C. Imbert, R. Monneau et H. Zidani s’intéressent à des équations de Hamilton-Jacobi
définies sur un réseau G ne comportant qu’un sommet O, faisant intervenir différents Hamiltoniens H i
sur chaque branche Ji , totalement indépendants l’un de l’autre, et une condition de transmission au
point O. Les Hamiltoniens H i sont supposés indépendants de la variable d’état, coercifs et fortement
convexes et les branches Ji semi-infinies.
Plus précisément, les équations étudiées sont de la forme
∂ t u(t, x) + H(x, ∂ x u(t, x)) = 0
u(0, x) = u0 (x)
où le Hamiltonien H est de la forme
§
H i (p)
H(x, p) =
maxi=1,··· ,N H i+ (pi )
dans (0, T ) × G ,
sur G ,
si x ∈ Ji \ {O} et p ∈ R,
si x = O et p = (p1 , · · · , pN ),
(1.3.3)
(1.3.4)
où le Hamiltonien H i+ peut être vu comme la partie décroissante du Hamiltonien H i . Notons que C.
Imbert et al utilisent la notation H i− au lieu H i+ , la signe − pour souligner le caractère décroissant. La
raison de notre choix d’utiliser H i+ plutôt que H i− est de rester cohérent avec les notations utilisées par
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
25
Y. Achdou et al dans [ACCT13] et déjà employées plus haut. Nous savons que H i est décroissant sur un
intervalle (−∞, p0i ] puis croissant sur l’intervalle [p0i , +∞), donc
§
H i (pi )
si pi ∈ (−∞, p0i ],
+
H i (pi ) =
i
minq∈R H i (q) = H i (p0 ) si pi ∈ [p0i , +∞).
L’équation (1.3.3) est la version non-stationnaire de l’équation proposée dans [ACCT13] et est clairement discontinue par rapport à la variable d’état.
L’espace des fonctions-test choisi pour définition les notions de solutions de viscosité dans (1.3.3) est
la version non-stationnaire de celui de [ACCT13]. Sous des hypothèses adaptées, les auteurs prouvent
que l’équation (1.3.3) admet une unique solution. Les méthodes développées pour prouver ces résultats sont très différentes de celles de [ACCT13]. L’existence est obtenue par la méthode de Perron.
L’unicité découle d’un principe de comparaison obtenu par des méthodes de type contrôle. Plus exactement, C. Imbert et al obtiennent le principe de comparaison en prouvant des principes de sous- et
sur-optimalité qui permettent de comparer les sous- et sur-solution de (1.3.3) avec une solution explicite de (1.3.3). Cette solution explicite est obtenue par une interprétation de l’équation (1.3.3) dans la
théorie du contrôle optimal. Les techniques de preuve utilisent fortement le fait que les Hamiltoniens
H i sont indépendants de la variable d’espace.
Enfin, il a été démontré par F. Camilli et C. Marchi dans [CM13] que les notions de solution de
viscosité proposées dans [ACCT13] et [IMZ13] sont équivalentes, et dans le cadre des eikonales, sont
aussi équivalentes à celle proposée dans [SC13].
Les résultats d’unicité proposés dans [ACCT13] et [IMZ13] ont ensuite été améliorés, dans le premier travail présenté dans ce mémoire d’une part, et dans [IM14a] d’autre part, par deux méthodes
différentes.
1.3.2.2
Réseaux multi-dimensionnels et espaces ramifiés
Le cas des multi-réseaux ou des espaces ramifiés généraux est plus compliqué que celui du réseau.
D’une part, comme les interfaces ne sont plus réduites à des points, les Hamiltoniens associées à ces
dernières peuvent mettre en jeu les dérivées tangentielles le long de l’interface. Dans le cas du contrôle,
on peut imaginer des dynamiques spécifiques tangentes à l’interface et des coûts associés : en un point
de l’interface, il faudra tenir compte des Hamiltoniens venant des différentes régions et du Hamiltonien
correspondant à ces dynamiques tangentes. Les avancées concernant les réseaux d’une part et celles
sur les équations de Hamilton-Jacobi sur des espaces stratifiées (voir § 1.2.2.1) d’autre part ont rendu
possible l’étude des cas multidimensionnels.
Dans [CSM13], F. Camilli, D. Schieborn et C. Marchi généralisent les résultats obtenus dans [SC13]
à des espaces ramifiés de dimensions supérieures. Plus exactement, les espaces ramifiés considérés
dans [CSM13] sont des unions disjointes de sous-variétés de RN de dimension N − 1 et N − 2. En
particulier, ce sont des multi-réseaux au sens de la définition donnée dans § 1.3.1. L’intérêt de se limiter
à ce genre d’espace ramifié est qu’en un point x d’une interface, on peut toujours définir le vecteur
νi (x), normal à l’interface en x et qui pointe vers la i-ème sous-variété de dimension N − 1 si celle-ci
est adjacente. Les auteurs utilisent fortement cette propriété pour généraliser les résultats de [SC13].
En particulier, c’est grâce à cette propriété que les auteurs peuvent utiliser un espace de fonctionstest analogue à celui introduit dans [SC13]. L’espace
de fonctions-test utilisé sur l’espace ramifié R, est
contenu dans l’espace C 0 (R)∩ ∩i∈I C 1 (M i ) , où Mi désigne la i-ème sous-variété de dimension N −1
de l’espace ramifié R, (R = ∪i∈I M i ), avec une condition de compatibilité supplémentaire aux points
singuliers portant sur les dérivées directionnelles dans les directions normales à l’interface νi (x). Bien
sûr, comme dans [SC13], seules des équations eikonales sont considérées.
1.3.3
1.3.3.1
Résultats obtenus par d’autres auteurs durant la préparation de la thèse
Les travaux de C. Imbert et R. Monneau
Ce paragraphe est consacré aux articles [IM14a, IM14b] de C. Imbert et R. Monneau qui ont été écrits
pendant le déroulement de la thèse. Bien que ces travaux concernent des aspects similaires à ceux que
j’ai étudiés, nous verrons que les techniques sont différentes.
26
CHAPITRE 1. INTRODUCTION
Dans [IM14a], C. Imbert et R. Monneau s’intéressent à des équations de Hamilton-Jacobi dépendant
du temps sur des réseaux. Plus précisément, l’objectif de cet article est de développer des méthodes
reposant sur des outils de la théorie des équations aux dérivées partielles et s’appliquant une classe
assez générale d’équations de Hamilton-Jacobi sur des réseaux. En particulier, les résultats de [IM14a]
s’appliquent à des Hamiltoniens quasi-convexes dans les arêtes et à des conditions de jonction très
générales.
Afin de simplifier la présentation des résultats, on se place dans un premier temps dans le cas où le
réseau ne comporte qu’un sommet et est défini de la manière suivante :
G = ∪i=1,··· ,N Ji
avec
Ji ∩ J j = {O} pour i 6= j,
et Ji est supposée isométrique à la demi-droite réelle [0, +∞), et O est appelé le point de jonction.
Avant de préciser le type d’équation étudié ici, donnons l’espace des fonctions-test qui sera utilisé pour
définir les notions de solution de viscosité
C 1 ((0, T ) × G ) = ϕ ∈ C 0 ((0, T ) × G ), la restriction de ϕ à (0, T ) × Ji est C 1 pour i = 1, · · · , N .
Pour une fonction ϕ ∈ C 1 ((0, T ) × G ), nous utiliserons la notation ∂ x ϕ(t, x) pour désigner
∂ x ϕ(t, x) =
§
∂i ϕ(t, x)
(∂1 ϕ(t, x), · · · , ∂N ϕ(t, x))
si x ∈ Ji \ {O},
si x = O,
où pour (t, x) ∈ (0, T ) × Ji et i = 1, · · · , N , la notation ∂i ϕ(t, x) désigne la différentielle usuelle de
ϕ|(0,T )×Ji par rapport à la variable d’état, qui est bien définie puisque ϕ ∈ C 1 ((0, T ) × G ).
Pour le réseau G , les équations de Hamilton-Jacobi que souhaitent étudier les auteurs se présentent
sous la forme
∂ t u(t, x) + H i (∂ x u(t, x)) = 0 dans (0, T ) × Ji \ {O},
∂ t u(t, x) + F (∂ x u(t, x)) = 0
dans (0, T ) × {O},
(1.3.5)
u(0, x) = u0 (x)
sur G ,
et les Hamiltoniens H i : R → R sont supposés satisfaire les propriétés suivantes :
. Continuité : H i ∈ C (R)
. Quasi-convexité : ∃ pi0 ∈ R tel que H i est décroissant sur (−∞, pi0 ] et croissant sur [pi0 , +∞)
. Coercivité :lim|q|→+∞ H i (q) = +∞.
La fonction de jonction F : RN → R est supposée satisfaire les propriétés suivantes :
. Continuité : F ∈ C (RN )
. Décroissance : F est décroissante dans chacune de ses variables.
Remarquons que la condition de jonction qui apparaît sur la deuxième ligne de l’équation (1.3.5) est
générale, car les hypothèses sont faites sur la fonction de jonction F sont minimales : la condition de
décroissance sur la fonction de jonction F est naturelle et apparaît comme nécessaire dans la preuve du
résultat d’existence par la méthode de Perron. Pour une fonction u : [0, T ) × G → R localement bornée,
nous rappelons que les notations u∗ et u∗ désignent les enveloppes semi-continues inférieurement et
supérieurement de u, définies par
u∗ (t, x) = lim inf u(s, y)
(s, y)→(t,x)
et
u∗ (t, x) = lim sup u(s, y).
(1.3.6)
(s, y)→(t,x)
Les auteurs introduisent deux notions de solution de viscosité. Dans la première, la condition de jonction est satisfaite exactement par toute fonction-test touchant par-dessous (resp. par-dessus) une sursolution (resp. sous-solution) au point de jonction O.
Définition 1.3.1 (Solution de viscosité).
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
27
. Une fonction u : [0, T ) × G → R localement bornée est une sous-solution (resp. sur-solution) de
(1.3.5) sur (0, T )×G si pour tout (t 0 , x 0 ) ∈ (0, T )×G et ϕ ∈ C 1 ((0, T )×G ) telle que u∗ −ϕ admet
un maximum (resp. u∗ − ϕ admet un minimum) local en (t 0 , x 0 ) ∈ (0, T ) on a
∂ t ϕ(t 0 , x 0 ) + H i (∂ x ϕ(t 0 , x 0 )) ≤ 0 (resp. ≥ 0)
∂ t ϕ(t 0 , x 0 ) + F (∂ x ϕ(t 0 , x 0 )) ≤ 0 (resp. ≥ 0)
si x 0 ∈ Ji∗ ,
si x 0 = O,
. Une fonction u : [0, T ) × G → R localement bornée est une sous-solution (resp. sur-solution) de
(1.3.5) sur [0, T ) × G si c’est une sous-solution (resp. sur-solution) de (1.3.5) sur (0, T ) × G et si de
plus
u∗ (0, x) ≤ u0 (x) (resp. u∗ (0, x) ≥ u0 (x)) pour tout x ∈ G .
. Une fonction u : [0, T ) × G → R localement bornée est une solution de (1.3.5) sur [0, T ) × G si c’est
une sous-solution et sur-solution de (1.3.5) sur [0, T ) × G .
Cependant, la notion de solution de viscosité proposée ci-dessus semble trop rigide pour donner
de bonnes propriétés de stabilité, particulièrement au point de jonction. C’est pourquoi C. Imbert et R.
Monneau introduisent une autre notion de solution de viscosité où la condition de jonction est relaxée.
Définition 1.3.2 (Solution de viscosité relaxée).
. Une fonction u : [0, T ) × G → R localement bornée est une sous-solution relaxée (resp. sur-solution
relaxée) de (1.3.5) sur (0, T ) × G si pour tout (t 0 , x 0 ) ∈ (0, T ) × G et ϕ ∈ C 1 ((0, T ) × G ) telle que
u∗ − ϕ admet un maximum (resp. u∗ − ϕ admet un minimum) local en (t 0 , x 0 ) ∈ (0, T ) on a
∂ t ϕ(t 0 , x 0 ) + H i (∂ x ϕ(t 0 , x 0 )) ≤ 0
(resp. ≥ 0),
quand x 0 ∈ Ji∗ , et
soit
soit
∂ t ϕ(t 0 , x 0 ) + F (∂ x ϕ(t 0 , x 0 )) ≤ 0 (resp. ≥ 0),
∂ t ϕ(t 0 , x 0 ) + H i (∂i ϕ(t 0 , x 0 )) ≤ 0 (resp. ≥ 0)
pour un certain i = 1, · · · , N ,
quand x 0 = O.
. Une fonction u : [0, T ) × G → R localement bornée est une sous-solution relaxée (resp. sur-solution
relaxée) de (1.3.5) sur [0, T ) × G si c’est une sous-solution relaxée (resp. sur-solution relaxée) de
(1.3.5) sur (0, T ) × G et si de plus
u∗ (0, x) ≤ u0 (x) (resp. u∗ (0, x) ≥ u0 (x))
pour tout x ∈ G .
. Une fonction u : [0, T ) × G → R localement bornée est une solution relaxée de (1.3.5) sur [0, T ) × G
si c’est une sous-solution relaxée et sur-solution relaxée de (1.3.5) sur [0, T ) × G .
Les auteurs prouvent que la notion de sous-solution relaxée (resp. sur-solution relaxée) est stable
par supremum (resp. infimum). Ainsi, ils parviennent à adapter la méthode de Perron pour prouver que
l’équation (1.3.5) admet une solution relaxée.
Théorème 1.3.1 (Existence). Si la condition initiale u0 est uniformément continue, l’équation (1.3.5)
admet une solution relaxée.
Afin de comprendre les liens entre les deux notions de solution évoquées ci-dessus, il nous faut
introduire une condition de jonction, dite condition de jonction à flux limité, qui joue un rôle particulier
dans l’étude proposée ici. Pour un paramètre A ∈ R ∪ {−∞}, appelé limiteur de flux, on définit une
fonction de jonction FA : RN → R par

‹
FA(p1 , · · · , pN ) = max A, max H i+ (pi ) ,
(1.3.7)
i=1,··· ,N
où le Hamiltonien H i+ est la partie décroissante de H i définie par
§
H i (q)
si q ≤ pi0 ,
H i+ (q) =
0
H i (pi ) si q ≥ pi0 .
(1.3.8)
28
CHAPITRE 1. INTRODUCTION
Rappelons que dans [IM14a], la notation H i− est utilisée au lieu de H i+ , mais que nous utilisons ici H i+
pour être cohérent avec les notations utilisées dans [ACCT13]. On parle alors de condition de jonction
à flux limité lorsque la fonction de jonction F qui apparaît dans (1.3.5) est égale à FA pour un certain
A ∈ R ∪ {−∞}. On peut remarquer que si on pose
A0 = max min H i (q),
i=1,··· ,N q∈R
alors pour tout A ∈ [−∞, A0 ] on a
FA(P1 , · · · , pN ) = max H i+ (pi ).
(1.3.9)
i=1,··· ,N
En d’autres termes, si A ≤ A0 , alors il n’y a pas de limitation du flux et on retrouve la condition de
jonction proposée dans l’article antérieur [IMZ13] écrit par les auteurs avec H. Zidani, voir (1.3.4), qui
est la version non stationnaire de la condition étudiée par Y. Achdou et al dans [ACCT13].
Lorsque la condition de jonction est une condition de jonction à flux limité (F = FA) les auteurs prouvent
que les deux notions de solution et de solution relaxée sont équivalentes :
Proposition 1.3.1 (équivalence entre les deux notions de solution lorsque F = FA).
Si F = FA pour un certain A ∈ R, alors les sous-solutions relaxées (resp. sur-solutions relaxées) de (1.3.5)
sur [0, T ) × G coïncident avec les sous-solutions (resp. sur-solutions) de (1.3.5) sur [0, T ) × G .
Par ailleurs, les auteurs prouvent que l’on peut toujours ramener l’étude de l’équation (1.3.5) à une
équation avec une condition de jonction à flux limité.
Proposition 1.3.2 (D’une condition de jonction générale à une condition de jonction à flux limité).
Pour toute fonction de jonction F, continue et décroissante par rapport à chacune des variables, il existe un
limiteur de flux A F ∈ [A0 , +∞) tel que toute sous-solution relaxée (resp. sur-solution relaxée) de (1.3.5)
sur [0, T ) × G soit une sous-solution (resp. sur-solution) de (1.3.5) sur [0, T ) × G avec la condition de
jonction Fe = FAF .
Un autre résultat important de [IM14a] est le principe de comparaison très général suivant, qui
permet de comparer les sous- et sur-solutions de (1.3.5), mais également les sous- et sur-solutions
relaxées de (1.3.5) avec une condition de jonction générale F .
Théorème 1.3.2 (Principe de comparaison).
Si la condition initiale u0 est uniformément continue, alors pour toute sous-solution (resp. sous-solution
relaxée) u de (1.3.5) et pour toute sur-solution (resp. sur-solution relaxée) v de (1.3.5), s’il existe C T > 0
tel que
u(t, x) ≤ C T (1 + d(O, x))
et
v(t, x) ≥ −C T (1 + d(O, x)),
∀(t, x) ∈ [0, T ) × G ,
où d( . , . ) est la distance géodésique sur la jonction G , alors
u≤v
dans
[0, T ) × G .
Grâce aux résultats qui précèdent, et principalement grâce aux Propositions 1.3.1 et 1.3.2, on sait
qu’il suffit de prouver le résultat de comparaison pour des sous- et sur-solutions (non relaxées) de (1.3.5)
avec une condition de jonction à flux limité FA, associée à un limiteur de flux A ∈ [A0 , +∞). Dans ce
dernier cas, la preuve du principe de comparaison proposée ici repose sur une adaptation de la méthode
de dédoublement des variables au contexte des réseaux. Une première idée est de remplacer le terme
|x− y|2
de pénalisation usuel " , qui est bien défini lorsque l’équation est posée sur un espace euclidien,
d(x, y)2
qui est plus adapté à la géométrie de G . Cependant, ce changement permet uniquement
par
"
de prendre en compte le changement de géométrie et ne permet pas de gérer les problèmes liés à la
discontinuité de l’équation (1.3.5) par rapport à la variable d’état. L’idée novatrice que proposent ici les
|x− y|2
y
auteurs est de remplacer le terme de pénalisation usuel " par un terme de la forme "G( "x , " ) , où
la fonction G est telle que l’inégalité
H( y, −∂ y G(x, y)) − H(x, ∂ x G(x, y)) ≤ 0,
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
29
est presque satisfaite pour x, y ∈ G proches de O, où H est le Hamiltonien associé à l’équation (1.3.5)
§
H i (p)
si x ∈ Ji∗ et p ∈ R,
H(x, p) =
FA(p1 , · · · , pN ) si x = O et (p1 , · · · , pN ) ∈ RN .
Si on arrive à construire une telle fonction G suffisamment régulière, il est alors facile de se convaincre
|x− y|2
que la méthode de dédoublement des variables où l’on remplace le terme de pénalisation usuel "
x y
par "G( " , " ) permet de conclure. Une telle fonction G est appelée fonction-test sommet. Pour mettre
en oeuvre cette preuve, la difficulté principale est évidement de construire cette fonction-test. L’idée de
base de la construction a été inspirée par des articles sur des lois de conservation scalaire avec fonctions
flux discontinues [GNPT07, AKR11]. Donnons en un aperçu dans le cas où N ≥ 2, c’est-à-dire quand le
réseau compte au moins deux branches. Si on note
GA = (p, λ) ∈ RN × R : H i (pi ) = FA(p) = λ ∀i = 1, · · · , N ,
et G 0 : G 2 → R la fonction définie par
G 0 (x, y) = sup
(p,λ)∈GA
pi x − p j y − λ
si x ∈ Ji et y ∈ J j ,
alors la fonction G 0 + A est presque un bon candidat pour G. Cependant, on ne peut pas prendre la
fonction G 0 + A telle quelle du fait qu’elle est singulière sur la diagonale D = {(x, y) ∈ G 2 : x = y}.
Dans le cas où les Hamiltoniens H i sont convexes, la fonction G est obtenue en régularisant la fonction
G 0 + A. Dans les cas où les Hamiltoniens H i sont seulement quasi-convexes, la fonction G est obtenue
en déformant les Hamiltoniens H i pour se ramener au cas convexe.
Les auteurs généralisent ensuite les résultats d’existence (Théorème 1.3.1) et d’unicité (Théorème 1.3.2)
prouvés pour l’équation (1.3.5) définie sur G . Les généralisations consistent d’une part à considérer des
réseaux comportant plusieurs sommets et d’autre part à autoriser des Hamiltoniens sur chaque branche
dépendant maintenant de la variable d’état et du temps. Les auteurs étendent l’approche décrite cidessus au cas des équations de Hamilton-Jacobi définies sur un réseau G quelconque :
∂ t u(t, x) + H e (t, x, ∂ x u(t, x)) = 0
∂ t u(t, x) + FAn (t, x, ∂ x u(t, x)) = 0
u(0, x) = u0 (x)
pour (t, x) ∈ (0, T ) × e∗ ,
pour (t, x) ∈ (0, T ) × {n},
sur G ,
(1.3.10)
où e désigne une arête de G , e∗ désigne l’arête e privée de son bord (ses sommets) et n désigne un
sommet de G . Dans (1.3.10), les Hamiltoniens H e (e parcourant l’ensemble des arêtes de G ) sont supposés continus et vérifient les mêmes propriétés que les Hamiltoniens H i (dans l’équation (1.3.5)) par
rapport à la variable du gradient. Par ailleurs, les conditions de jonction qui apparaissent dans l’équation (1.3.10) sont des conditions de jonction à flux limité, avec des limiteurs de flux An possiblements
différents sur chaque sommet n de G .
Dans [IM14b] C. Imbert et R. Monneau montrent que la théorie qu’ils ont développée dans [IM14a]
pour des équations de Hamilton-Jacobi posées sur des réseaux s’adapte bien au cas des multi-jonctions.
Par multi-jonction, on entend un ensemble J ⊂ RN construit comme une union finie de demi-hyperplans
Ji qui partagent le même bord Γ , que nous appellerons l’interface. L’interface Γ est donc un sous-espace
vectoriel de RN de dimension N − 2. En d’autres terme, une multi-jonction J se décompose comme suit
J = ∪i=1,··· ,N Ji
et
Ji ∩ J j = Γ pour i 6= j,
où la i-ème branche Ji peut se ramener à RN −2 ×R+ par un changement de coordonnées. Un point x ∈ Ji
s’écrit x = (x 0 , x i ) : x 0 ∈ RN −2 est la composante de x dans la direction tangentielle à Γ et x i ∈ R+ est
la composante de x dans la direction normale à Γ qui pointe vers Ji . Les auteurs introduisent un espace
de fonctions-test analogue à celui qu’ils avaient introduit dans [IM14a],
C 1 ((0, T ) × J) = ϕ ∈ C 0 ((0, T ) × J), la restriction de ϕ à (0, T ) × Ji est C 1 pour i = 1, · · · , N .
Pour une fonction ϕ ∈ C 1 ((0, T ) × J), nous utiliserons la notation ∂ x ϕ(t, x) pour désigner
∂ x 0 ϕ(t, x), ∂ x i ϕ(t, x)
si x ∈ Ji \ Γ ,
∂ x ϕ(t, x) =
∂ x 0 ϕ(t, x), (∂ x 1 ϕ(t, x), · · · , ∂ x N ϕ(t, x))
si x ∈ Γ ,
où
30
CHAPITRE 1. INTRODUCTION
. pour (t, x) ∈ (0, T ) × Ji et i = 1, · · · , N , la notation ∂ x i ϕ(t, x) désigne la dérivée partielle par
rapport à la variable x i de ϕ|(0,T )×Ji , qui est bien définie puisque ϕ ∈ C 1 ((0, T ) × J)
. pour (t, x) ∈ (0, T ) × Ji , avec x = (x 0 , x i ), la notation ∂ x 0 ϕ(t, x) désigne la différentielle de
l’application x 0 7→ ϕ(t, (x 0 , x i )).
Les équations de Hamilton-Jacobi considérées dans [IM14b] se présentent sous la forme
∂ t u(t, x) + H i (∂ x u(t, x)) = 0
∂ t u(t, x) + F (∂ x u(t, x)) = 0
u(0, x) = u0 (x)
dans (0, T ) × Ji \ Γ ,
dans (0, T ) × Γ ,
sur J.
(1.3.11)
Les Hamiltoniens H i : RN −1 → R sont supposés satisfaire les propriétés suivantes :
. Continuité : H i ∈ C (RN −1 ) ;
. Quasi-convexité : pour tout λ ∈ R l’ensemble p ∈ RN −1 : H i (p) ≤ λ est convexe ;
. Coercivité :lim|q|→+∞ H i (q) = +∞.
La fonction d’interface F : RN −2 × RN → R est supposée satisfaire les propriétés suivantes :
. Continuité : F ∈ C (RN −2 × RN ) ;
. Décroissance : pour tout i ; l’application pi 7−→ F (p0 , p1 , · · · , pi , · · · , pn ) est décroissante ;
. Quasi-convexité : pour tout λ ∈ R l’ensemble p ∈ RN −2 × RN : F (p) ≤ λ est convexe.
Comme dans [IM14a], les auteurs introduisent deux notions de solution de viscosité pour (1.3.11) :
une notion de solution équivalente à celle donnée dans la Définition 1.3.1 et une notion de solution relaxée équivalente à celle donnée dans la Définition 1.3.2. Comme dans [IM14a], les auteurs obtiennent
l’existence d’une solution relaxée en appliquant la méthode de Perron.
Théorème 1.3.3 (Existence). Si la condition initiale u0 est uniformément continue, l’équation (1.3.5)
admet une solution relaxée.
De manière analogue à [IM14a], les conditions d’interface à flux limité jouent un rôle prépondérant :
Pour une fonction continue et quasi-convexe A : RN −2 → R, dite fonction limiteur de flux, on définit la
fonction d’interface FA : RN −2 × RN par
‹

FA(p0 , p1 , · · · , pN ) = max A(p0 ), max H i+ (p0 , pi ) ,
i=1,··· ,N
où le Hamiltonien H i+ : RN −1 → R est défini par
H i+ (p0 , pi ) =
§
H i (p0 , pi )
H i (p0 , π0i (p0 ))
si pi ≤ π0i (p0 ),
si pi ≥ π0i (p0 ).
(1.3.12)
On parle alors de condition d’interface à flux limité lorsque il existe une fonction limiteur de flux A :
RN −2 → R telle que la fonction d’interface F qui apparaît dans (1.3.5) est égale à FA. Les conditions
d’interface à flux limité jouent le même rôle que dans [IM14a]. On peut remarquer que lorsque la
fonction A est telle que pour tout p0 ∈ RN −2 on a A(p0 ) ≤ A0 (p0 ) = maxi=1,··· ,N minq∈R H i (p0 , q), alors la
fonction d’interface FA s’écrit
FA(p0 , p1 , · · · , pN ) = max H i+ (p0 , pi ).
i=1,··· ,N
(1.3.13)
C. Imbert et al prouvent des résultats analogues aux Propositions 1.3.1 et 1.3.2. En particulier, comme
dans [IM14a], la preuve d’un principe de comparaison général pour l’équation (1.3.5) se ramène à la
preuve d’un principe de comparaison pour des sous- et sur-solutions de (1.3.5) au sens de la Définition
1.3.1 dans le cas d’une condition d’interface à flux limité (F = FA). Les auteurs prouvent un principe de
comparaison analogue à celui énoncé Théorème 1.3.2 :
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
31
Théorème 1.3.4 (Principe de comparaison).
Si la condition initiale u0 est uniformément continue, alors pour toute sous-solution (resp. sous-solution
relaxée) u de (1.3.5) et pour toute sur-solution (resp. sur-solution relaxée) v de (1.3.5), si il existe C T > 0
tel que
u(t, x) ≤ C T (1 + d(0, x))
et
v(t, x) ≥ −C T (1 + d(0, x)),
∀(t, x) ∈ [0, T ) × J,
où d( . , . ) est la distance géodésique sur J, alors
u≤v
dans
[0, T ) × J.
La technique utilisée pour prouver ce résultat est similaire à celle développée pour prouver le Théorème 1.3.2 dans [IM14a]. Il s’agit d’une méthode de dédoublement des variables où l’on remplace le
y
|x− y|2
terme de pénalisation usuel " par un terme de la forme "G( "x , " ), où la fonction G est construite
pour pouvoir comparer les Hamiltoniens issus de différentes branches.
1.3.3.2
Autres travaux
Dans [HZ15], C. Hermosilla et H. Zidani étudient des problèmes de contrôle optimal à horizon infini
avec contrainte d’état pour lesquels la variable d’état doit appartenir à un ensemble fermé K de RN à
structure stratifiée, les strates Mi étant des sous-variétés disjointes de RN telles que si Mi ∩M j 6= ; alors
Mi ⊂ M j . La classe des ensembles K est plus générale que celle des espaces ramifiés définis dans le
paragraphe § 1.3.1, car des sous-variétés de RN de dimension N sont permises. En particulier, l’ensemble
K n’est pas nécessairement d’intérieur vide et pourrait être un espace stratifié du type de ceux étudiés
dans les articles [BH07, BBC13, BBC14, BC14, RZ13, RSZ14] évoqués dans le paragraphe § 1.2.2. À
la différence d’autres travaux dans lesquels la dynamique est définie de manière indépendante dans
chaque sous-variété, la dynamique est décrite comme une fonction continue sur RN ×A, où A est un sousensemble compact de Rm . Une hypothèse de contrôlabilité est faite pour approcher des trajectoires qui
oscillent infiniment entre plusieurs strates dans un intervalle de temps donné : cette hypothèse est écrite
en terme d’ensembles atteignables dans une strate. Elle permet seulement d’assurer que la fonction
valeur est semi-continue inférieurement. Le but de [HZ15] est de caractériser la fonction valeur comme
l’unique sur-solution bilatérale semi-continue inférieurement d’une équation de Hamilton-Jacobi posée
sur K . La formulation, adaptée à la géométrie, met en jeu une condition de sur-solution relativement
classique et des conditions de sous-solution moins habituelles impliquant des Hamiltoniens tangentiels
sur chaque sous-variété, et n’est pas sans rappeler celles proposées dans [BH07, BC14]. C. Hermosilla
et al font aussi l’hypothèse que le coût instantané est une fonction définie sur RN × A et localement
lipschitzienne par rapport à la variable d’état. L’existence est déduite d’un principe de programmation
dynamique et l’unicité découle d’un principe de comparaison prouvé par des arguments de type contrôle,
en passant par des principes de sur- et sous-optimalité.
Le cadre géométrique des articles [GHN14, Nak14] est beaucoup plus large que celui des espaces
ramifiés, puisqu’il s’agit d’espaces métriques généraux (χ, d). Ils englobent les espaces ramifiés, mais
aussi des espaces métriques structurellement beaucoup plus complexes tels que certains ensembles
fractals.
Dans [GHN14] Y. Giga, N. Hamamuki et A. Nakayasu étudient des équations eikonales sur des espaces
métriques généraux. Plus précisément, les auteurs considèrent des équations eikonales de la forme
| Du(x) |= f (x)
dans Ω,
(1.3.14)
où Ω est un ouvert d’un espace métrique (χ, d). Le fait de se placer sur un espace métrique (χ, d)
général présente une difficulté particulière, puisque la différentielle Dξ d’une fonction ξ : χ → R
n’est généralement pas définie. Cependant, les auteurs arrivent à traiter le cas des équations eikonales,
comme (1.3.14), puisqu’ils parviennent à donner un sens au module de la différentielle | Dξ | d’une
fonction absolument continue ξ : χ → R. Cette notion faible permet en particulier de construire un
espace de trajectoires admissibles régulières A x (R, Ω) qui jouera un rôle important dans les définitions
de sous- et sur-solution de viscosité de (1.3.14). En effet, une idée importante utilisée ici par les auteurs
consiste à composer les fonctions u : Ω → R, candidates à être sous- ou sur-solution de viscosité, par des
32
CHAPITRE 1. INTRODUCTION
trajectoires admissibles ξ ∈ A x (R, Ω) judicieusement choisies, ce qui permet de se ramener à l’étude de
fonctions définies sur R. Ainsi, Y. Giga et al proposent des notions de sous- et sur-solution de viscosité
de (1.3.14) sur Ω qui reposent sur des choix appropriés de trajectoires admissibles ξ ∈ A x (R, Ω) et
des outils d’analyse réelle. Les auteurs prouvent un résultat d’existence en montrant que la fonction
valeur du problème de contrôle optimal naturellement associé à l’équation (1.3.14) est bien solution de
viscosité de cette dernière. Ils prouvent également un principe de comparaison en adaptant la méthode
de H. Ishii, développée dans [Ish87b].
Dans [Nak14] A. Nakayasu généralise les résultats de [GHN14] au cas des équations évolutives, de
la forme
∂ t u(t, x) + H(x, | Du(t, x) |) = 0 dans (0, T ) × χ,
(1.3.15)
u(0, x) = u0 (x)
sur χ.
sur des espaces métriques généraux. Notons que l’équation (1.3.15) dépend du module de la différentielle | Du |. Comme on l’a déjà signalé, ceci est nécessaire pour pouvoir travailler sur un espace
métrique (χ, d) général. L’auteur propose des notions de sous- et sur-solution de viscosité pour l’équation (1.3.15) qui reposent comme dans [GHN14] sur l’introduction d’un espace de fonction admissible.
Comme dans [GHN14] l’auteur obtient l’existence et l’unicité d’une solution pour l’équation (1.3.15).
Le principe de comparaison est obtenu par des méthodes différentes de celles de [GHN14], reposant
sur des arguments de type contrôle.
Un autre article profond sur le sujet est celui de L. Ambrosio et J. Feng [AF14] dans lequel les auteurs
considèrent une classe assez générale d’équations de Hamilton-Jacobi du premier ordre sur des espaces
métriques géodésiques complets. En particulier, les équations étudiées dans [GHN14, Nak14] entrent
dans le champ d’application de [AF14]. L’approche proposée ici est très différente de celle suggérée
dans [GHN14, Nak14], reposant entièrement sur des outils d’analyse pour les espaces métriques. Voir
[AGS08, Chapitre 1] pour une introduction aux outils de bases de l’analyse sur des espaces métriques.
Les notions de sous- et sur-solutions proposées ici sont proches des définitions usuelles et reposent sur
des choix d’espaces de fonctions-test appropriés à la structure d’espace métrique. On peut signaler que
les espaces des fonctions-test qui interviennent dans les définitions de sous-solution et sur-solution sont
différents. En particulier, les définitions de sous- et sur-solution sont non-symétriques. L. Ambrosio et
al prouvent des résultats d’existence en passant par des formules de représentation et des principes de
comparaison en adaptant la méthode de dédoublement des variables à leur contexte.
Dans la lignée de ces travaux on peut citer l’article de F. Camilli, R. Capitanelli et C. Marchi [CCM14]
dans lequel ils étudient l’équation eikonale sur le triangle de Sierpinski S, qui est un espace métrique
particulier de type fractale. Les auteurs considèrent d’abord l’équation eikonale sur les pré-fractales
S n , qui sont des réseaux. Pour ces équations, F. Camilli et al utilisent la notion de solution de viscosité
introduite par D. Schieborn et F. Camilli dans [SC13]. D’après l’étude faite dans [SC13] on sait que
l’équation eikonale sur le pré-fractale S n admet une unique solution un pour laquelle on dispose d’une
formule de représentation. Pour l’équation eikonale sur le triangle de Sierpinski S les auteurs utilisent
la notion de solution de viscosité introduite par Y. Giga, N. Hamamuki et A. Nakayasu dans [GHN14].
D’après [GHN14], on sait que l’équation eikonale sur le triangle de Sierpinski admet une unique solution
u pour laquelle on dispose d’une formule de représentation. Les auteurs prouvent que (à extraction
d’une sous-suite près) la suite des solutions un de l’équation eikonale sur le pré-fractale S n converge
uniformément vers la solution u de l’équation eikonale sur le triangle de Sierpinski S. Enfin, on peut
signaler que les auteurs proposent également une approximation discrète de la solution u de l’équation
eikonale sur le triangle de Sierpinski S, obtenue en considérant des équations eikonales définies sur
l’ensemble des sommets des pré-fractales S n .
1.3.4
Résumé des Chapitres 2 et 3
Ce paragraphe est consacré à la présentation des travaux qui apparaissent dans la Partie II de ce manuscrit et qui s’inscrivent dans le développement de la théorie des équations de Hamilton-Jacobi-Bellman
discontinues sur des réseaux ou plus généralement des multi-réseaux. Cette partie se décompose en
deux chapitres, le Chapitre 2 qui traite du cas du réseau et le Chapitre 3 qui traite le cas d’une multijonction.
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
1.3.4.1
33
Résumé du Chapitre 2
Le Chapitre 2 est un travail en commun avec Y. Achdou et N. Tchou qui a fait l’objet d’un article publié
dans ESAIM : Control, Optimisation and Calculus of Variations [AOT15b]. Dans [AOT15b], nous nous
intéressons à des problèmes de contrôle optimal à horizon infini pour lesquels les trajectoires restent
sur un réseau et pour lesquels on autorise différentes dynamiques et différents coûts instantanés sur
chaque arête du réseau. L’objectif de ce travail est de montrer que l’on peut adapter les arguments de
la théorie du contrôle optimal introduits par G. Barles, A. Briani et E. Chasseigne dans [BBC13,BBC14]
pour obtenir une preuve simple d’un principe de comparaison. Dans un premier temps, on considère
le cas particulier où le réseau G dans Rd est constitué de N branches infinies fermées, avec un seul
sommet. Plus précisément, on suppose que le réseau G se décompose sous la forme
G = ∪i=1,··· ,N Ji
où
J i = R+ e i ,
les vecteurs ei étant supposés unitaires et deux à deux disjoints. Nous noterons O l’origine qui correspond au point de jonction. Afin de pouvoir exposer clairement le problème considéré ici, commençons
par préciser les données du problème ainsi que les différentes hypothèses. Les données du problème
sont les dynamiques f i : Ji × Ai → R et les coûts instantanés `i : Ji × Ai → R associés à chaque branche,
où Ai est un sous-ensemble compact d’un espace métrique A. On fait l’hypothèse non restrictive que les
ensembles Ai sont deux à deux disjoints. Les hypothèses de régularité faites sur les fonctions f i et `i
sont relativement classiques. On suppose également que les ensembles
FLi (x) = {( f i (x, a)ei , `i (x, a)) : a ∈ Ai } ,
(1.3.16)
sont non vides, fermés et convexes. Enfin, on suppose qu’il existe un réel δ > 0 tel que pour tout
i = 1, · · · , N
[−δ, δ] ⊂ { f i (O, a) : A ∈ Ai } .
(1.3.17)
Ceci est une hypothèse de contrôlabilité localisée au point de jonction O qui assure l’existence de
contrôles permettant de s’approcher ou de s’éloigner de la jonction dans un voisinage de cette dernière.
Pour définir la dynamique et le coût instantané sur G entier, nous introduisons l’ensemble
M = (x, a); x ∈ G , a ∈ Ai si x ∈ Ji \{O}, et a ∈ ∪Ni=1 Ai si x = O .
On définit alors la fonction dynamique f sur M par
§
f i (x, a)ei
∀(x, a) ∈ M ,
f (x, a) =
f i (O, a)ei
si x ∈ Ji \{O},
si x = O et a ∈ Ai ,
et le coût instantané ` sur M par
∀(x, a) ∈ M ,
`(x, a) =
§
`i (x, a)
`i (O, a)
si x ∈ Ji \{O},
si x = O et a ∈ Ai .
Les fonctions f et ` sont bien définies et continues sur M puisque l’on a supposé que les ensembles
Ai sont disjoints. Ainsi, l’ensemble des trajectoires admissibles partant de x ∈ G de notre problème de
contrôle optimal est donné par


∞
+
y x ∈ Lip(R+ ; G ),

 ( y x , α) ∈ LLoc (R ; M ) :
Z t
Tx =
,
+
y x (t) = x +
f ( y x (s), α(s))ds sur R 

0
et la fonction valeur s’écrit
v(x) =
Z
∞
inf
( y x ,α)∈T x
`( y x (t), α(t))e−λt d t,
0
et λ est une constante positive strictement. Un premier résultat est un principe de programmation
dynamique qui implique que la fonction valeur est continue.
34
CHAPITRE 1. INTRODUCTION
Proposition 1.3.3. La fonction valeur v est bornée et continue sur G .
Avant de donner l’équation de Hamilton-Jacobi-Bellman qui caractérise la fonction valeur v, commençons par introduire l’espace de fonction-test.
Définition 1.3.3. Une fonction φ : G → R est une fonction-test admissible si
. φ est continue sur G
. pour tout j, j = 1, . . . , N , φ|J j ∈ C 1 (J j ).
On note R(G ) l’ensemble des fonctions-test admissibles.
L’équation de Hamilton-Jacobi-Bellman candidate pour caractériser la fonction valeur v est l’équation
dans Ji \ {O},
λu(t, x) + H i (x, ddux i (x)) = 0
(1.3.18)
du
λu(t, x) + HO ( ddu
(x),
·
·
·
,
(x))
=
0
sur {0},
x1
d xN
où le Hamiltonien H i : Ji × R → R est donné par
H i (x, p) = max (− f i (x, a)p − `i (x, a)) ,
a∈Ai
et le Hamiltonien HO : RN → R est défini par
HO (p1 , . . . , pN ) = max H i+ (pi ),
i=1,...,N
où le Hamiltonien
H i+
: R → R est défini par
H i+ (p) =
a∈Ai
max
(− f i (O, a)p − `i (O, a)).
t.q. f i (O,a)≥0
(1.3.19)
Le Hamiltonien H i+ défini par (1.3.19) diffère de H i , car au lieu de prendre tous les contrôles a ∈ Ai , on
ne considère que ceux qui correspondent à des directions qui permettent de rentrer dans la branche Ji .
C’est pour cette raison que nous avons fait le choix de la notation H i+ comme on l’a déjà écrit. Notons
que H i+ est la partie décroissante du Hamiltonien H i . En particulier, la condition de jonction proposée
ici est similaire à celle proposée dans [ACCT13] et dans [IM14a] lorsque le limiteur de flux est tel que
A ≤ A0 , voir (1.3.9). La définition de solution de viscosité est la suivante :
Définition 1.3.4 (Solution de viscosité).
. Une fonction semi-continue supérieurement (resp. inférieurement) u : G → R est une sous-solution
(resp. sur-solution) de (1.3.18) sur G si pour tout x ∈ G et pour tout φ ∈ R(G ) tels que u − φ a
un maximum (resp. minimum) local en x, alors
dφ
λu(x) + H i (x, d x i (x)) ≤ 0
λu(O) +
dφ
dφ
HO ( d x 1 (O), . . . , d x N
(resp. ≥ 0)
(O)) ≤ 0
si x ∈ Ji \{O},
(resp. ≤ 0)
si x = O.
. Une fonction continue u : G → R est une solution de (1.3.18) sur G si c’est une sous-solution et une
sur-solution de (1.3.18) sur G .
L’existence d’une solution de (1.3.18) est obtenue de la même manière que dans [ACCT13] en
passant par l’équation
λu(x) + sup {−Du(x, ζ) − ξ} = 0 dans G ,
(1.3.20)
e
(ζ,ξ)∈ f`(x)
qui correspond à un problème de contrôle optimal relaxé. Dans l’équation (1.3.20) ci-dessus, l’ensemble
e
f`(x)
est défini par
e
f`(x)
=

t n → 0+ ,

Z tn



1
∃( y x,n , αn )n∈N ,

lim
f ( y x,n (t), αn (t))d t = η,
(η, µ) ∈ Rd × R : ( y x,n , αn ) ∈ T x , t.q. n→∞ t n 0
Z tn


∃(t n )n∈N

lim 1

`( y x,n (t), αn (t))d t = µ

n→∞
tn 0











.
(1.3.21)
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
35
e
et pour φ ∈ R(G ) et (ζ, ξ) ∈ f`(x)
la notation Dφ(x, ζ) désigne
¨ dφ
si x ∈ Ji \ {O},
d x i (x)ζ.ei
Dφ(x, ζ) =
dφ
si x = 0 et ζ ∈ Rei .
d x i (O)ζ.ei
On propose une définition de solution de viscosité pour l’équation (1.3.20) analogue à celle proposée
pour l’équation (1.3.18).
Définition 1.3.5 (Solution de viscosité de l’équation relaxée).
. Une fonction semi-continue supérieurement (resp. inférieurement) u : G → R est une sous-solution
(resp. sur-solution) de (1.3.20) sur G si pour tout x ∈ G et pour tout φ ∈ R(G ) tels que u − φ a
un maximum (resp. minimum) local en x, alors
λu(x) +
sup
{−Dφ(x, ζ) − ξ} ≤ 0 (resp. ≥ 0).
e
(ζ,ξ)∈ f`(x)
. Une fonction continue u : G → R est une solution de (1.3.20) sur G si c’est une sous-solution et une
sur-solution de (1.3.20) sur G .
La méthode suivie pour prouver l’existence d’une solution de (1.3.18) est la suivante : on prouve
d’abord l’existence pour l’équation (1.3.20). En effet, on peut montrer que la fonction valeur v est
solution de cette équation en passant par le principe de programmation dynamique. Alors, on conclut
en montrant que les solutions de l’équation (1.3.20) coïncident avec les solutions de l’équation (1.3.18).
Un des arguments clefs pour prouver la correspondance entre ces deux équations est que l’on est capable
e
d’exprimer l’ensemble f`(x)
en fonction des ensembles F L i (x), définis par (1.3.16), et des ensembles
+
F L i (O) définis par
FL+i (O) = FLi (O) ∩ (R+ ei × R).
Plus précisément, on prouve que
e
f`(x)
=
FL (x)
Si
si x ∈ Ji \ {O},
¦
Š©
S €
+
j6=i FL j (O) ∩ ({0} × R)
i=1,...,N co FLi (O) ∪
si x = O.
Finalement, on obtient le résultat d’existence
Théorème 1.3.5 (Existence). La fonction valeur v est une solution bornée de l’équation (1.3.18) sur G .
L’équation de Hamilton-Jacobi tangentielle au sommet O va jouer un rôle essentiel :
λu(O) + HOT = 0
sur {O},
(1.3.22)
où
T
HOT = max HO,i
,
i=1,··· ,N
avec
T
HO,i
=
a∈Ai
max
−`i (O, a).
t.q. f i (O,a)=0
T
Le Hamiltonien HO,i
est construit de la même manière que le Hamiltonien H i+ , défini par (1.3.19), mais
en ne gardant cette fois que les contrôles a ∈ Ai qui permettent de rester sur le point de jonction O.
On prouve le résultat suivant :
Lemme 1.3.1. Il existe un voisinage de O dans G de la forme B(O, r) ∩ G , avec r > 0, dans lequel toute
sous-solution bornée de (1.3.18) est lipschitzienne.
Ce lemme est une conséquence de l’hypothèse de contrôlabilité (1.3.17). En particulier, les soussolutions de (1.3.18) sont lipschitziennes que dans un voisinage du point de jonction O car l’hypothèse
de contrôlabilité est localisée au point de jonction O. Une conséquence importante du Lemme 1.3.1 est
le résultat suivant :
36
CHAPITRE 1. INTRODUCTION
Lemme 1.3.2. Toute sous-solution bornée de (1.3.18) est une sous-solution de (1.3.22).
En effet, si u est une sous-solution de (1.3.18), le Lemme 1.3.1 nous assure l’existence d’une fonctiontest ϕ ∈ R(S ) qui touche u par dessus au point de jonction O. Alors, en écrivant l’inégalité de soussolution en O avec cette fonction-test et en remarquant que HOT ≤ HO on obtient le résultat.
Pour les sur-solutions la situation est plus compliquée. On a une alternative qui a été inspirée par les
travaux de G. Barles, A. Briani et E. Chasseigne dans [BBC13, BBC14].
Théorème 1.3.6 (Alternative). Soit r > 0 donné par le Lemme 1.3.1. Soit w : G → R une sur-solution de
viscosité de (1.3.18), bornée inférieurement par −C(|x| + 1) où C est un réel positif. Alors l’une des deux
propriétés ci-dessous est vérifiée :
[A] Il existe ηk une suite de réels positifs tels que limk→+∞ ηk = η > 0, un indice i ∈ {1, . . . , N } et une
suite x k ∈ Ji \ {O}, telle que limk→+∞ x k = O, limk→+∞ w(x k ) = w(O) et pour tout k, il existe une
loi de contrôle αki telle que la trajectoire correspondante y x k (s) ∈ Ji ∩ B(O, r) pour tout s ∈ [0, ηk ] et
w(x k ) ≥
Z
ηk
`i ( y x k (s), αki (s))e−λs ds + w( y x k (ηk ))e−ληk ,
0
[B]
λw(O) + HOT ≥ 0.
La propriété [A] est une propriété de sur-optimalité localisé le long d’une suite de trajectoires admissibles restant dans une des branches du réseau (au moins pour des temps court) et la propriété [B]
est une inégalité de sur-solution pour l’équation tangentielle (1.3.22).
On prouve le principe de comparaison suivant.
Théorème 1.3.7 (Principe de comparaison). Soit u : G → R une sous-solution bornée de (1.3.18), et
w : G → R une sur-solution bornée de (1.3.18). Alors u ≤ w sur G .
La preuve de ce principe de comparaison combine des arguments d’équation aux dérivées partielles
et des arguments de la théorie du contrôle optimal. Tout d’abord, le Hamiltonien étant continu en
dehors du point de jonction, c’est à dire en x ∈ G \ {O}, on peut conclure par la méthode standard
de dédoublement des variables en ces points. Au point de jonction O, si la sur-solution w satisfait
l’alternative [B] dans le Théorème 1.3.6, on peut conclure très simplement en écrivant que u et w sont
respectivement sous- et sur-solution de l’équation tangentielle (1.3.22) ; si la sur-solution w satisfait
l’alternative [A], on utilise des arguments de contrôle optimal.
Corollaire 1.3.1. La fonction valeur v est l’unique solution de viscosité bornée de (1.3.18).
On prouve également des résultats de stabilité pour les sous- et sur-solutions de (1.3.18) lorsqu’on
perturbe les données du problème de contrôle optimal :
si on se donne des suites ( f i" )">0 et (`"i )">0 de dynamiques et de coûts instantanés telles que
f i" , `"i −→ f i0 , `0i
"→0
localement uniformément sur Ji × Ai ,
et si les fonctions f i" , `"i satisfont les hypothèses avec des constantes uniformes en ", alors la suite de
fonctions valeur v " converge localement uniformément vers la fonction valeur v 0 .
L’étude proposée ici dans le cas d’un réseau à un seul sommet s’étend très bien au cas d’un réseau
G quelconque. En effet, les raisonnements qui apparaissent dans les différentes preuves sont tous des
raisonnements locaux.
Enfin, on peut également prouver que les preuves de principe de comparaison proposées ici s’adaptent
bien à des conditions de jonction plus générales, impliquant un limiteur de flux A ∈ R, de la forme
‹

du
du
(x), · · · ,
(x)) = 0 sur {0}.
λu(t, x) + max A, HO (
d x1
d xN
qui sont du même type que dans [IM14a].
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
1.3.4.2
37
Résumé du Chapitre 3
Dans le Chapitre 3, on cherche à généraliser les méthodes développées dans le Chapitre 2 au cas des
multi-jonctions. Ce travail fait l’objet d’une prépublication [Oud14]. Une multi-jonction S dans Rd
est une union de N demi-hyperplans fermés Pi , i = 1, · · · , N , s’intersectant en un espace affine Γ de
dimension d − 2. Sans perte de généralité on peut se placer dans le cas d’une multi-jonction de R3 de
sorte que la multi-jonction S peut s’écrire sous la forme
S = ∪i=1,··· ,N Pi
où
Pi = Re0 × R+ ei ,
et les vecteurs ei sont des vecteurs unitaires deux à deux distincts tels que ei .e0 = 0 pour tout i =
1, · · · , N . Dans ce cas, l’interface Γ s’écrit
Γ = Re0 .
On considère donc un problème de contrôle optimal à horizon infini sur la multi-jonction S pour lequel
on autorise différentes dynamiques f i : Pi ×Ai → Re0 ×Rei et différents coûts instantanés `i : Pi ×Ai →
R dans chaque région Pi de la multi-jonction S . Les hypothèses sur les données sont essentiellement les
mêmes que celles faites dans le cas des réseaux. Seules les hypothèses de contrôlabilité seront précisées
dans la suite. Comme précédemment, le problème de contrôle optimal n’est pas bien défini tant que
l’on n’a pas précisé la dynamique f et le coût instantané ` aux points de l’interface x ∈ Γ . On définit
l’ensemble
M = (x, a); x ∈ S , a ∈ Ai si x ∈ Pi \Γ , et a ∈ ∪Ni=1 Ai si x ∈ Γ .
La dynamique f sur M est ainsi définie par
∀(x, a) ∈ M ,
f (x, a) =
§
f i (x, a)
f i (x, a)
si x ∈ Pi \Γ ,
si x ∈ Γ et a ∈ Ai ,
`(x, a) =
§
`i (x, a)
`i (x, a)
si x ∈ Pi \Γ ,
si x ∈ Γ et a ∈ Ai .
et la fonction ` sur M par
∀(x, a) ∈ M ,
Ceci nous permet de définir l’ensemble des trajectoires admissible partants de x ∈ S


y x ∈ Lip(R+ ; S ),


Z t
∞
,
T x = ( y x , α) ∈ LLoc
(R+ ; M ) : +
f ( y x (s), α(s))ds sur R 

y x (t) = x +
0
et la fonction valeur
v(x) =
Z
∞
inf
( y x ,α)∈T x
`( y x (t), α(t))e−λt d t.
0
Le cadre étant maintenant bien défini, il nous faut préciser l’hypothèse de contrôlabilité avec laquelle
nous travaillons. Dans un premier temps nous allons faire une hypothèse de contrôlabilité forte, puis
l’affaiblir.
Premier cas : avec de la contrôlabilité forte près de l’interface Γ
On fait l’hypothèse de contrôlabilité forte suivante : ∃δ > 0 tel que ∀i = 1, · · · , N et ∀x ∈ Γ ,
B(0, δ) ∩ (Re0 × Rei ) ⊂ { f i (x, a) : a ∈ Ai } .
(1.3.23)
Comme dans le cas du réseau, la fonction valeur v satisfait un principe de programmation dynamique,
et l’hypothèse de contrôlabilité forte (1.3.23) nous permet de prouver la continuité de la fonction valeur.
Proposition 1.3.4. La fonction valeur v est bornée et continue sur S .
L’espace de fonctions-test choisi ici est l’analogue de celui qui a été introduit dans le cas des réseaux.
Définition 1.3.6. Une fonction φ : S → R est une fonction-test admissible si
38
CHAPITRE 1. INTRODUCTION
. φ est continue sur S
. pour tout j, j = 1, . . . , N , φ|P j ∈ C 1 (P j ).
On note R(S ) l’ensemble des fonctions-test admissibles.
Par ailleurs, on s’attend à ce que la fonction valeur soit caractérisée par l’équation de HamiltonJacobi-Bellman ci-dessous :
si x ∈ Pi \Γ ,
λu(x) + H i (x, D ϕ|Pi (x)) = 0
(1.3.24)
λu(x) + HΓ x, D ϕ|P1 (x), . . . , D ϕ|PN (x) = 0
si x ∈ Γ ,
où le Hamiltonien H i : Pi × (Re0 × Rei ) → R est donné par
H i (x, p) = max (− f i (x, a.)p − `i (x, a)) ,
a∈Ai
et le Hamiltonien HΓ : Γ ×
Q
i=1,...,N (Re0
× Rei ) → R est défini par
HΓ (x, p1 , . . . , pN ) = max H i+ (x, pi ),
i=1,...,N
où H i+ : Pi × (Re0 × Rei ) → R est défini par
H i+ (x, p) =
a∈Ai
max
(−p. f i (x, a) − `i (x, a)).
t.q. f i (x,a).ei ≥0
La condition d’interface proposée ici est similaire à celle proposée dans [IM14b] dés que le limiteur
de flux vérifie A(p0 ) ≤ A0 (p0 ), voir (1.3.13). La définition de solution de viscosité de(1.3.24) est la
suivante :
Définition 1.3.7 (Solution de viscosité).
. Une fonction semi-continue supérieurement (resp. inférieurement) u : S → R est une sous-solution
(resp. sur-solution) de (1.3.24) sur S si pour tout x ∈ S et pour tout φ ∈ R(S ) tels que u − φ a
un maximum (resp. minimum) local en x, alors
si x ∈ Pi \Γ ,
λu(x) + H i (x, D ϕ|Pi (x)) ≤ 0 (resp. ≥
0) si x ∈ Γ .
λu(x) + HΓ x, D ϕ|P1 (x), . . . , D ϕ|PN (x) ≤ 0 (resp. ≤ 0)
. Une fonction continue u : S → R est une solution de (1.3.24) sur S si c’est une sous-solution et une
sur-solution de (1.3.24) sur S .
Comme pour le cas du réseau, l’existence est démontrée en considérant l’équation de HamiltonJacobi-Bellman correspondant à un problème de contrôle optimal relaxé
λu(x) +
sup
{−Du(x, ζ) − ξ} = 0
dans S ,
(1.3.25)
e
(ζ,ξ)∈ f`(x)
e
où l’ensemble f`(x)
est défini de manière analogue à (1.3.21) et où pour ϕ ∈ R(S )
D ϕ|Pi (x).ζ si x ∈ Pi \ Γ ,
Dϕ(x, ζ) =
D ϕ|Pi (x).ζ si x ∈ Γ et ζ ∈ Re0 × Rei .
La définition de solution de viscosité proposée pour l’équation (1.3.25) est l’analogue à la définition
1.3.5. Comme conséquence du principe de programmation dynamique, la fonction valeur v est une
e
solution de l’équation (1.3.25). Par ailleurs on peut montrer que l’ensemble f`(x)
peut se décomposer
sous la forme
FLi (x)
¦
Š© si x ∈ Pi \ Γ ,
e
S €
f`(x)
= S
+
si x ∈ Γ ,
i=1,...,N co FLi (x) ∪
j6=i FL j (x) ∩ (Re0 × R)
où pour x ∈ Γ l’ensemble FL+i (x) est défini par
FL+i (x) = FLi (x) ∩ (Re0 × R+ ei ) × R.
Cette décomposition permet de prouver que les solutions des équations (1.3.25) et (1.3.24) coïncident.
Finalement on obtient le résultat suivant :
1.3. ÉQUATIONS DE HAMILTON-JACOBI SUR DES RÉSEAUX
39
Théorème 1.3.8 (Existence). La fonction valeur v est une solution de viscosité bornée de l’équation
(1.3.24) sur S .
Comme dans le cas du réseau, il y a une équation tangentielle associée à l’équation (1.3.24) qui va
jouer un rôle clef dans la preuve du principe de comparaison. Il s’agit de l’équation
λu(x) + HΓT (x, D(ϕ|Γ )(x)) = 0
sur Γ ,
(1.3.26)
où le Hamiltonien HΓT : Γ × Re0 → R est défini par
HΓT (x, p) = max HΓT,i (x, p),
i=1,...,N
avec le Hamiltonien HΓT,i : Γ × Re0 → R qui est défini par
HΓT,i (x, p) =
a∈Ai
max
(− f i (x, a).p − `i (x, a)).
t.q. f i (x,a).ei =0
On prouve le résultat analogue au Lemme 1.3.1 suivant :
Lemme 1.3.3. Il existe un voisinage de Γ dans S , de la forme B(Γ , r) ∩ S avec r > 0 (où B(Γ , r) désigne
l’ensemble des points de Rd dont la distance à Γ est inférieure à r), dans lequel toute sous-solution bornée
u de (1.3.24) est lipschitzienne.
Une conséquence importante de ce lemme est le résultat ci-dessous, qui est l’analogue du Lemme
1.3.2.
Lemme 1.3.4. Soit u une sous-solution de viscosité bornée de (1.3.24) et φ : Γ → R une fonction C 1 .
Alors, pour tout x̄ ∈ Γ tel que u − φ a un maximum local en x̄ sur Γ ,
λu(x̄) + HΓT (x̄, D(φ|Γ )(x̄)) ≤ 0.
(1.3.27)
La preuve de ce lemme demande un peu plus de travail que celle du Lemme 1.3.2. Cependant,
l’idée de la preuve reste la même. On utilise le fait que la fonction u est lipschitzienne pour prolonger
la fonction φ : Γ → R au domaine S tout entier en une fonction φ̃ appartenant à R(S ) qui reste au
dessus de u. Alors on en déduit l’inégalité (1.3.27) en remarquant que HΓT ≤ HΓ .
Pour les sur-solutions de (1.3.24), on obtient une alternative analogue au Théorème 1.3.6.
Théorème 1.3.9 (Une alternative). Soit r > 0 donné par le Lemme 1.3.1. Soit w : S → R une sursolution de viscosité de (1.3.24), bornée inférieurement par −C(|x| + 1), où C est un réel positif. Soit
φ ∈ R(S ) et x̄ ∈ Γ tels que w − φ ait un minimum local en x̄. Alors l’une des deux propositions ci-dessous
est vérifiée :
[A] Il existe ηk une suite de réels positifs tels que limk→+∞ ηk = η > 0, un indice i ∈ {1, . . . , N } et une
suite x k ∈ Pi \ Γ , telle que limk→+∞ x k = x̄, limk→+∞ w(x k ) = w(x̄) et pour tout k, il existe une loi
de contrôle αki telle que la trajectoire correspondante y x k (s) ∈ Pi ∩ B(Γ , r) pour tout s ∈ [0, ηk ] et
w(x k ) ≥
Z
ηk
`i ( y x k (s), αki (s))e−λs ds + w( y x k (ηk ))e−ληk .
0
[B]
λw(x̄) + HΓT (x̄, D(φ|Γ )(x̄)) ≥ 0.
En suivant les lignes de la démonstration du Théorème 1.4 dans 1.3.10 dans [BBC13] on parvient
à montrer le principe de comparaison suivant :
Théorème 1.3.10 (Principe de comparaison). Soit u : S → R une sous-solution bornée de (1.3.24), et
w : S → R une sur-solution bornée de (1.3.24). Alors u ≤ w sur S .
40
CHAPITRE 1. INTRODUCTION
La preuve du Théorème 1.3.10 est plus technique que celle du Théorème 1.3.7. En effet, dans le
cas du réseau, l’alternative du Théorème 1.3.6 nous fournissait directement des informations sur la
sur-solution w. Ici, pour pouvoir appliquer l’alternative du Théorème 1.3.9, il nous faut choisir judicieusement une fonction-test φ ∈ R(S ) telle que w − φ ait un minimum local en x̄ ∈ Γ . Pour construire
la fonction-test φ, nous utilisons la technique développée par G. Barles, A. Briani et E. Chasseigne
dans [BBC13]. Cette technique consiste à régulariser la sous-solution u (par convolution dans la direction tangentielle à l’interface) de manière à obtenir une fonction u" qui soit de classe C 1 sur Γ et qui
reste presque une sous-solution de (1.3.24). En faisant un raisonnement classique de dédoublement
des variables, on parvient à construire une fonction-test φ ∈ R(S ) permettant d’appliquer l’alternative
dans le Théorème 1.3.9 et telle que D(φ|Γ )(x̄) = D(u" |Γ )(x̄). Avec ce choix de fonction-test φ, on arrive
à conclure quasiment de la même manière que dans la preuve du Théorème 1.3.6.
Du Théorème 1.3.10, on déduit le corollaire suivant :
Corollaire 1.3.2. La fonction valeur v est l’unique solution de viscosité bornée de (1.3.24).
Deuxième cas : avec de la contrôlabilité normale dans un voisinage de l’interface Γ
On affaiblit hypothèse de contrôlabilité de la façon suivante : ∃δ > 0, tel que pour tout i = 1, · · · , N et
x ∈ Γ,
[−δ, δ] ⊂ { f i (x, a).ei : a ∈ Ai } .
(1.3.28)
On parle de contrôlabilité normale. Cette hypothèse de contrôlabilité est similaire à celle faite par G.
Barles, A. Briani et E. Chasseigne dans [BBC14]. Comme dans le cas de la contrôlabilité forte, on souhaite montrer que la fonction valeur v est l’unique solution de viscosité bornée de l’équation (1.3.24).
Cependant, sous l’hypothèse de contrôlabilité normale, on ne peut plus déduire du principe de programmation dynamique que la fonction valeur v est continue. Il va donc falloir travailler dans un premier
temps avec une notion de solution discontinue.
Définition 1.3.8 (Solution discontinue). Une fonction u : S → R bornée est une solution discontinue de
(1.3.24) dans S si u? est une sous-solution de (1.3.24) dans S et u? est une sur-solution de (1.3.24) dans
S (on rappelle que u? et u? désignent les enveloppes semi-continues inférieurement et supérieurement de
u, définies par (1.3.6)).
Théorème 1.3.11 (Existence). La fonction valeur v est une solution discontinue de l’équation (1.3.24)
sur S .
De même, l’hypothèse de contrôlabilité normale ne conduit pas au caractère lipschitzien des soussolutions dans un voisinage de l’interface Γ . Le Lemme 1.3.4 n’est plus satisfait que par les sous-solutions
de (1.3.24) qui sont lipschitziennes dans un voisinage de Γ . En revanche, l’alternative du Théorème
1.3.9 reste valable avec l’hypothèse de contrôlabilité normale.
Les sous-solutions n’étant plus lipschitziennes dans un voisinage de Γ , on ne peut plus prouver directement un principe de comparaison global. On prouve d’abord un principe de comparaison local :
Théorème 1.3.12. Soit u une sous-solution semi-continue supérieurement, bornée de (1.3.24) dans S et
w une sur-solution semi-continue inférieurement, bornée de (1.3.24) dans S . Il existe R > 0 tel que pour
tout y0 ∈ Γ fixé, si on pose Q = B( y0 , R) ∩ S ,
k (u − v)+ k L ∞ (Q) ≤k (u − v)+ k L ∞ (∂ Q) .
Comme on vient de le signaler, la sous-solution u n’est a priori plus lipschitzienne dans un voisinage
de Γ . Pour remédier à ce problème, on régularise la sous-solution u par la sup-convolution
(
2p )
|z0 − x 0 |2
+α
,
si x = x 0 e0 + x i ei ∈ Pi .
uα (x) := max u(z0 e0 + x i ei ) −
z0 ∈R
α2
On montre alors que la fonction uα est lipschitzienne dans l’ensemble Q et reste presque une soussolution de (1.3.24). On est donc ramené à manipuler une sous-solution lipschitzienne près de Γ et on
peut donc conclure en suivant les lignes de la preuve du principe de comparaison Théorème 1.3.10.
Remarquons que la preuve du caractère lipschitzien de la fonction uα utilise fortement la propriété
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
41
de contrôlabilité normale. En effet, la sup-convolution conduit à une fonction lipschitzienne dans la
direction tangentielle à l’interface, et la contrôlabilité normale permet alors de prouver le caractère
lipschitzien dans la direction normale. Le principe de comparaison global du Théorème 1.3.10 se déduit
facilement du principe de comparaison local du Théorème 1.3.12. Finalement, on déduit du principe
de comparaison global Théorème 1.3.10 que la fonction valeur v est continue et est l’unique solution
de (1.3.24).
Corollaire 1.3.3. La fonction valeur v est continue et est l’unique solution de viscosité bornée de (1.3.24).
Remarquons que, sous l’hypothèse de contrôlabilité forte comme sous l’hypothèse de contrôlabilité
normale, on peut étendre le cadre en ajoutant une dynamique f0 et un coût instantané `0 sur l’interface.
On obtient alors une condition d’interface de la forme
λu(x) + max H0 (x, D (u|Γ ) (x)), HΓ x, D u|P1 (x), . . . , D u|PN (x) = 0 sur Γ ,
où le Hamiltonien H0 : Γ × Re0 7−→ R est donné par
H0 (x, p0 ) = max (− f0 (x, a).p0 − `0 (x, a)) .
a∈A0
On retrouve alors des conditions d’interface similaire à celle proposée par C. Imbert et R. Monneau
dans [IM14b]. Finalement, les notions de solution de viscosité proposées dans [IM14b] et [Oud14]
coïncident.
1.4
Homogénéisation et perturbations singulières
On s’intéresse à des problèmes d’homogénéisation ou de perturbations singulières mettant en jeu des
équations de Hamilton-Jacobi discontinues par rapport à la variable d’état.
1.4.1
Problématique
La théorie de l’homogénéisation s’intéresse à l’étude de phénomènes se produisant dans des milieux
qui apparaissent comme homogènes à l’échelle macroscopique mais qui présentent une structure organisée ou non à l’échelle microscopique. Plus précisément, la théorie de l’homogénéisation permet de
définir rigoureusement le passage à la limite d’une équation comportant une ou plusieurs échelles microscopiques vers une équation limite régissant le comportement macroscopique. Si la micro-structure
du milieu étudié présente des discontinuités, l’intuition veut que l’équation à l’échelle microscopique
présente des discontinuités par rapport à la variable d’état. Dans la Figure 1.5 ci-dessous, on présente
des exemples de micro-structures hétérogènes.
(a)
(b)
FIGURE 1.5 – Exemples de micro-structures hétérogènes
Depuis les articles précurseurs de P-L. Lions, G. Papanicolaou et S. Varadhan [LPV] et de C. Evans
[Eva89], la théorie de l’homogénéisation pour les équations de Hamilton-Jacobi s’est considérablement
développée. Les techniques maintenant classiques de la théorie de l’homogénéisation utilisent très fortement des résultats d’existence mais surtout des propriétés de stabilité et d’unicité forte (i.e. des principes
de comparaison). On a vu dans la partie § 1.2 que la question de l’unicité dans le cadre des équations
42
CHAPITRE 1. INTRODUCTION
de Hamilton-Jacobi discontinues présente des difficultés. Ainsi, c’est seulement assez récemment que
la communauté de la théorie des solutions de viscosité a commencé à s’intéresser à la question de l’homogénéisation d’équations de Hamilton-Jacobi avec des Hamiltoniens discontinus.
Considérons une équation de Hamilton-Jacobi dont le Hamiltonien fait intervenir une échelle rapide
dans la variable d’état, par exemple une dépendance en x/", où " est un petit paramètre destiné à tendre
vers 0. Il est possible que d’éventuelles discontinuités dans la variable d’état soient présentes à l’échelle
microscopique mais n’apparaissent plus dans l’équation effective après passage à la limite. On s’attend à
ce genre de situations pour des Hamiltoniens périodiques en x/". En revanche, le milieu à homogénéiser
peut présenter des discontinuités à l’échelle macroscopique en plus des variations rapides déjà évoquées.
On peut par exemple penser à un milieu périodique sauf dans de petites régions comportant des défauts.
Ce genre de problèmes a fait l’objet du cours de P-L. Lions au Collège de France en 2014, [Lio14], où
la notion d’homogénéisation précisée a été introduite. On peut également envisager des problèmes
d’homogénéisation impliquant des structures à singularité géométrique telles que des réseaux ou des
espaces ramifiés : si la géométrie ne dépend pas de ", ces problèmes peuvent être considérés comme des
problèmes d’homogénéisation précisée, car le passage à la limite quand " tend vers 0 doit tenir compte
de la singularité géométrique qui ne disparaît pas.
Un autre cas de figure est celui de problèmes d’homogénéisation pour lesquels une structure ramifiée
R" ⊂ RN se densifie quand le paramètre " tends vers 0, et “tend” vers un RN ou une partie de RN de
dimension N à la limite. Dans ce cas, la variable d’état de l’équation de départ prend ses valeurs dans
une structure de dimension strictement inférieure à celle de l’espace où est posée l’équation limite.
Nous verrons aussi des exemples de problèmes de perturbations singulières pour lesquels les équations
de Hamilton-Jacobi qui dépendent de " sont continues et telles que l’équation effective est discontinue.
Nous pensons en particulier à des problèmes de perturbations singulières posés dans une famille Ω"
d’ouverts régulier de RN convergeant vers un domaine ramifié (de dimension inférieur à N ) quand "
tend vers 0. Dans ce cas, la variable d’état de l’équation limite prend ses valeurs dans une structure
de dimension strictement inférieure à celle de l’espace où est posée l’équation de départ. Les questions
de perturbations singulières des équations de Hamilton-Jacobi font l’objet d’une vaste littérature, à
partir des travaux de A. Bensoussan [Ben88]. On peut distinguer les approches utilisant la théorie du
contrôle optimal, voir par exemple [GL99, AG00], des approches basées sur la théorie des EDP, voir par
exemple [AB02, AB03, ABM08].
1.4.2
Résultats obtenus par d’autres auteurs durant la préparation de la thèse
Nous commençons par présenter quelques problèmes d’homogénéisation pour lesquels l’équation de
Hamilton-Jacobi de départ est discontinue et l’équation effective est continue.
Dans [FR14], N. Forcadel et Z. Rao s’intéressent à des problèmes de perturbations singulières dans
le cadre du contrôle optimal sur des domaines stratifiés. Plus précisément, on partitionne l’espace R2
en damier, comme dans la Figure 1.5 (b). On note Ω1 l’ouvert constitué par les cases blanches et Ω2
l’ouvert constitué par les cases foncées. On note Γ l’interface entre Ω1 et Ω2 (∂ Ω1 = ∂ Ω2 = Γ ). Ainsi,
pour tout " > 0, on peut écrire
Rd × R2 = "Ω1 ∪ "Ω2 ∪ "Γ .
Les auteurs s’intéressent à des problèmes de contrôle optimal dans Rd × R2 pour lesquels on impose
différentes dynamiques bi" : Rd × "Ωi × A → Rd × R2 sur chaque sous-domaine Rd × "Ωi de Rd × R2 ,
i = 1, 2. Les dynamiques bi" sont de la forme
f (x, y, a)
"
,
bi (x, y, a) = 1
" g i (x, y, a)
où " > 0, A est compact et f et g i sont des fonctions continues, lipschitziennes par rapport à la variable d’état. La première variable x ∈ Rd est une variable lente alors que la seconde y ∈ R2 est une
variable rapide. Aux points de l’interface (x, y) ∈ Rd × "Γ , on autorise des combinaisons convexes des
dynamiques issues de chaque côté de l’interface. Cela permet de définir un ensemble de trajectoires
"
admissibles S[t,T
(x, y) et de considérer le problème de Mayer
]
¦
©
"
v " (t, x, y) = inf ϕ(X (T ), Y (T )) : (X (·), Y (·) ∈ S[t,T
(x,
y)
,
]
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
43
où ϕ est une fonction lipschitzienne. D’après les résultats Z. Rao et H Zidani dans [RZ13], la fonction
valeur v " est l’unique solution de l’équation de Hamilton-Jacobi discontinue
−∂ t v " (t, x, y) + H i (x, y, Dx v " (t, x, y), 1" D y v " (t, x, y)) = 0
−∂ t v " (t, x, y) + H E (x, y, Dx v " (t, x, y), 1" D y v " (t, x, y)) = 0
v " (T, x, y) = ϕ(x, y)
dans (O, T ) × Rd × "Ωi ,
dans (O, T ) × Rd × "Γ ,
dans Rd × R2 ,
où H i est le Hamiltonien donné par
H i (x, y, p, q) = max (− f (x, y, a).p − g i (x, y, a).q) ,
a∈A
et où H E est le Hamiltonien essentiel défini dans (1.2.9). Les auteurs prouvent que la suite des fonctions
valeurs v " converge vers l’unique solution v de l’équation effective
−∂ t v(t, x) + H(x, Dx v(t, x)) = 0
v(T, x) = inf y∈R2 ϕ(x, y)
dans (O, T ) × Rd ,
dans Rd ,
où le Hamiltonien H(x, p) est défini comme étant l’unique constante λ pour laquelle le problème de
cellule
H i (x, y, p, D y w(t, y)) = λ
dans (O, T ) × Rd × Ωi ,
H E (x, y, p, D y w(t, y)) = λ dans (O, T ) × Rd × Γ ,
(1.4.1)
w(t, ·) est periodique.
admet une solution. Remarquons que le problème effectif dépend continûment de la variable d’état. La
résolution du problème de cellule est obtenue en considèrant le problème perturbé
δwδ (t, y) + H i (x, y, p, D y wδ (t, y)) = 0
δwδ (t, y) + H E (x, y, p, D y wδ (t, y)) = 0
wδ (t, ·) est periodique.
dans (O, T ) × Rd × Ωi ,
dans (O, T ) × Rd × Γ ,
(1.4.2)
Pour pouvoir résoudre le problème de cellule, N. Forcadel et Z. Rao prouvent des résultats d’existence et
de stabilité pour l’équation discontinue (1.4.2). Pour assurer l’unicité de la constante λ pour laquelle le
problème de cellule (1.4.1) admet une solution, ils prouvent un principe de comparaison valable pour
les sous- et sur-solutions lipschitziennes. La preuve du résultat de convergence se fait par la méthode
de la fonction-test perturbée de Evans [Eva89].
Dans [BBCT14], G. Barles, A. Briani, E. Chasseigne et N. Tchou s’intéressent à des problèmes d’homogénéisation dans le cadre d’un problème de contrôle optimal où l’on autorise différentes dynamiques
et coûts instantanés dans des régions complémentaires de l’espace RN . Plus exactement, on suppose que
l’espace RN se décompose de la façon suivante
R N = Ω 1 ∪ Ω2 ∪ H
où
Ω1 ∩ Ω 2 = ;
et
∂ Ω1 = ∂ Ω2 = H .
On suppose de plus que les ensembles Ωi sont ZN -périodique, c’est-à-dire que pour tout x ∈ Ωi et
k ∈ ZN on a x + k ∈ Ωi , et que l’interface H est régulière (W 2,∞ ). Par exemple, la partition du plan
proposée Figure 1.5 (a) rentre dans le cadre proposé ici, alors que celle proposée Figure 1.5 (b) n’est
pas assez régulière pour être prise en compte ici. Les auteurs s’intéressent à l’équation de HamiltonJacobi-Bellman
λu" (x) + H i (x, "x , Du" (x))= 0
dans "Ωi ,
mini=1,2 H i (x, "x , Du" (x)) ≤ 0 sur "H ,
(1.4.3)
x
maxi=1,2 H i (x, " , Du" (x)) ≥ 0 sur "H ,
où le Hamiltonien H i est donné par
H i (x, y, p) = sup (−bi (x, y, a).p − `i (x, y, a)) ,
a∈Ai
les fonctions bi et `i étant supposées ZN -périodiques par rapport à la variable y. D’après [BBC13,
BBC14], on sait qu’il n’y a pas unicité pour (1.4.3). En revanche, on connaît la solution maximale U"+
44
CHAPITRE 1. INTRODUCTION
et la solution minimale U"− de (1.4.3). Les auteurs souhaitent déterminer des équations effectives qui
caractérisent les limites des suites de fonctions U"+ et U"− . On a vu dans la partie § 1.2 que la fonction
U"− et U"+ peuvent s’interpréter comme étant les fonctions valeur de problèmes de contrôle optimal avec
deux ensembles de trajectoires admissibles, l’ensemble de trajectoires admissibles étant plus restreint
pour U"+ et constitué de trajectoires dites régulières. Par ailleurs, la fonction U"− peut être caractérisée
comme étant l’unique solution de (1.4.3) qui satisfait aussi l’équation tangentielle
λu(x) + H T (x,
x
, DH u(x)) ≤ 0
"
sur H ,
où le Hamiltonien H T apparaît dans l’équation (1.2.5). Ainsi, la détermination de l’équation effective
et l’étude de la convergence de la solution minimale U"− peuvent se faire en utilisant des techniques
d’équations aux dérivées partielles. Grâce aux nombreux résultat d’existence, de stabilité et de comparaison prouvés dans [BBC13, BBC14], les auteurs parviennent sans trop de difficultés à adapter les
méthodes standard à ce problème. Ils prouvent que la suite de fonctions U"− converge vers l’unique
solution U − de l’équation effective
−
λu(x) + H (x, Du(x)) = 0
dans RN ,
(1.4.4)
−
où le Hamiltonien H est obtenu à partir d’un problème de cellule.
L’étude de la suite U"+ est plus compliquée puisque ces dernières ne peuvent être caractérisées par
des équations aux dérivées partielles. Les auteurs parviennent à étudier la convergence de la suite U"+
en passant par des arguments de contrôle optimal. En particulier, le problème de cellule qui permet
+
de déterminer le Hamiltonien effectif H s’écrit par l’intermédiaire d’une formule de représentation
associée à un problème de contrôle optimal. Plus précisément, les auteurs prouvent que pour tout
+
x, p ∈ RN , il existe une unique constante H (x, p) pour laquelle il existe une fonction V + périodique
N
telle que ∀τ ≥ 0 et y0 ∈ R
Z τ
€
Š
+
+
+
˜
`(x, p, Y y (t), a(t)) + H (x, p) d t + V (Y y (τ)) ,
(1.4.5)
V ( y0 ) =
inf
0
r eg
(Y y0 ,a)∈T y0
0
0
où T yr eg est l’ensemble des trajectoires admissibles partant de y0 qui intervient dans le problème de
0
contrôle optimal associé à U1+ , et ˜` s’exprime à partir des fonctions bi et `i . La méthode suivie pour
prouver ce résultat est relativement proche de la méthode classiquement utilisée dans la théorie des
équations aux dérivées partielles. Les auteurs considèrent la suite V ρ+ des sous-solutions maximales
d’une équation de Hamilton-Jacobi-Bellman perturbée. D’après [BBC13,BBC14], les fonctions V ρ+ sont
bien définies et on dispose de formules de représentation pour ces dernières. Les auteurs parviennent
alors à conclure en exploitant une propriété de stabilité de l’équation perturbée et les formules de
représentation. La preuve du résultat de convergence est relativement technique car on ne dispose pas
de principe de comparaison pour (1.4.3). Cependant, on sait que toute sous-solution lipschitzienne de
(1.4.3) est inférieure à U"+ . Cette observation permet de montrer de manière presque standard que
U + = lim"→0 U"+ est sur-solution de l’équation effective
+
λu(x) + H (x, Du(x)) = 0
dans RN .
(1.4.6)
La preuve que U + est une sous-solution est plus complexe et utilise fortement une formule de représentation pour le correcteur V + donné par le problème de cellule (1.4.5). Finalement, les auteurs parviennent à prouver que la suite des fonctions U"+ converge vers l’unique solution de l’équation effective
(1.4.6). On peut signaler que les deux équations de Hamilton-Jacobi effectives (1.4.4) et (1.4.6) sont
continues.
La question de l’homogénéisation a également été abordée par C. Imbert et R. Monneau dans leur
article [IM14a] que nous avons déjà évoqué dans la partie § 1.3.3.1. En effet, comme application de leur
principe de comparaison les auteurs se proposent d’étudier un problème d’homogénéisation dans Rd
impliquant des réseaux. Plus précisément, C. Imbert et al considèrent la famille de réseaux N" , plongés
dans Rd , dont l’ensemble des sommets est donné par "Zd et l’ensemble des arêtes, que nous noterons
E" , est constitué des segments permettant de relier les sommets voisins. Dans la Figure 1.6 ci-dessous,
on représente le réseau N" dans le cas de la dimension 2.
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
"
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
"
45
FIGURE 1.6 – Illustration du réseau N" dans le cas de la dimension 2
Si {e1 , · · · , ed } désigne la base canonique de Rd , nous notons E"i l’ensemble des arêtes de N" qui
sont colinéaires au vecteur de base ei . Avec cette notation, l’équation de Hamilton-Jacobi à l’échelle "
considérée ici est donnée par
∂ t u" (t, x) + H i (∂ x u" (t, x)) = 0
∂ t u" (t, x) + FA(∂ x u" (t, x)) = 0
u" (0, x) = u0 (x)
si t > 0, x ∈ e∗ pour un certain e ∈ E"i ,
si t > 0, x ∈ "Zd ,
sur N" ,
(1.4.7)
où pour une arête e ∈ E" la notation e∗ désigne l’intérieur de cette arête, où les Hamiltoniens H i vérifient les mêmes hypothèses que les Hamiltoniens H i qui interviennent dans l’équation (1.3.5) et où la
fonction de jonction FA est définie comme dans (1.3.7) à partir d’un limiteur de flux A ∈ R donné et des
Hamiltoniens H i+ (définis par(1.3.8)), i = 1, · · · , d. D’après l’étude proposée dans [IM14a], l’équation
(1.4.7) admet une unique solution de viscosité u" , au sens de la définition 1.3.1. Les auteurs prouvent
que la suite des fonctions u" converge quand " tend vers 0 vers l’unique solution u0 d’une équation
effective dans Rd de la forme
∂ t u0 (t, x) + H(∂ x u0 (t, x)) = 0
u0 (0, x) = u0 (x)
dans (0, +∞) × Rd ,
sur x ∈ Rd ,
avec un Hamiltonien effectif H. Les techniques utilisées pour déterminer le Hamiltonien effectif H et
pour prouver le résultat de convergence sont relativement standards et utilisent fortement le fait que
l’on dispose d’un principe de comparaison pour l’équation de Hamilton-Jacobi sur le réseau (1.4.7).
Dans [FS14b] N. Forcadel et W. Salazar cherchent à obtenir un modèle de trafic routier macroscopique (décrivant l’évolution de la densité de véhicules) à partir d’un modèle microscopique (décrivant le
comportement de chaque véhicule). Le modèle microscopique de départ est un modèle introduit par M.
Bando et al dans [BHN+ 95] dans lequel les véhicules se meuvent sur une ligne droite et où l’évolution
du j-ème véhicule est régie par l’équation différentielle ordinaire
€
Š
U j00 (t) = a j Vj U j+1 (t) − U j (t) − U j0 (t) .
(1.4.8)
Ici U j désigne la position du j-ème véhicule, U j0 sa vitesse et U j00 son accélération. Le coefficient a j est la
sensibilité du j-ème conducteur et Vj (·) est la fonction de vitesse optimale. Le modèle proposé par M.
Bando et al autorise de choisir des sensibilités a j et des fonctions de vitesse optimale Vj (·) différentes
pour chaque conducteur. Ici, afin de pouvoir appliquer les techniques de l’homogénéisation périodique
les auteurs font l’hypothèse de périodicité supplémentaire suivante : il existe n0 ≥ 1 tel que
a j+n0 = a j
et
Vj+n0 (·) = Vj (·)
pour tout j ∈ Z.
(1.4.9)
En exploitant les équations (1.4.8) et (1.4.9) les auteurs parviennent à ramener l’étude du modèle décrit
ci-dessus à l’étude du système d’équations différentielles d’inconnue ((u j ) j∈Z , (ξ j ) j∈Z )
∂ t u j (t, x)
∂ t ξ j (t, x)
=
=
α ξ j (t, x) − u j (t, x) ,
(a j − α) u j (t, x) − ξ j (t, x) +
aj
α
Vj u j+1 (t, x) − u j (t, x) ,
(1.4.10)
46
CHAPITRE 1. INTRODUCTION
où α ∈ R est connu, avec la condition de périodicité
u j+n0 (t, x) = u j (t, x + 1)
et
ξ j+n0 (t, x) = ξ j (t, x + 1),
j
)
n0
et
ξ j (t, x) = ξ0 (x +
(1.4.11)
et la condition initiale
u j (t, x) = u0 (x +
j
).
n0
(1.4.12)
N. Forcadel et al prouvent que le système (1.4.10)-(1.4.11)-(1.4.12) admet une unique solution
((u j ) j∈Z , (ξ j ) j∈Z ), et que les fonctions remises à l’échelle u"j (t, x) = "u j ( "t "x ) et ξ"j (t, x) = "ξ j ( "t , "x )
convergent quand " tend vers 0 vers l’unique solution de viscosité u0 d’une équation effective de la
forme
∂ t u0 (t, x) = F (∂ x u0 (t, x)) dans (0, +∞) × R,
u0 (0, x)
= u0 (x)
sur R.
Cette équation régit l’évolution macroscopique du trafic. Le Hamiltonien effectif F est obtenu en considérant un problème intermédiaire qui peut être vu comme un problème de cellule et la preuve du
résultat de convergence utilise la méthode de la fonction-test perturbée de Evans.
Nous présentons maintenant l’étude de quelques problèmes d’homogénéisation précisée ou de perturbation singulière pour lesquels l’équation de Hamilton-Jacobi de départ est continue et l’équation
effective est discontinue.
Dans [AT15], Y. Achdou et N. Tchou considèrent une famille d’ouverts en forme d’étoile Ω" , qui
peuvent être vu comme des épaississements d’épaisseur " d’un réseau G , et qui convergent vers le réseau
G quand " tend vers 0. Y. Achdou et N. Tchou considèrent alors la suite des fonctions valeur v" associées
aux problèmes de contrôle optimal à horizon infini pour lesquels les trajectoires sont contraintes à rester
dans les ouverts Ω" . Ils souhaitent montrer que la suite des fonctions valeur v" converge quand " tend
vers 0 vers l’unique solution v d’une équation de Hamilton-Jacobi-Bellman sur G . Commençons par
définir plus précisément le réseau G et les ensembles Ω" . On suppose que G est défini par
G = ∪i=1,··· ,N Ji
où
J i = R+ e i ,
les vecteurs ei étant supposés unitaires et deux à deux distincts. On note O le point de jonction et ei⊥
un vecteur unitaire perpendiculaire à ei , pour i = 1, · · · , N . On considère alors Ω un ouvert étoilé par
rapport à O, à bord régulier. On suppose de plus qu’il existe un réel r0 > 0 tel que
Ω \ Wr0 = ∪i=1,··· ,N Si
où
Si = {z ∈ R2 : z.ei > r0 et − 1 < z.ei⊥ < 1}.
où
Wr0 = z ∈ R2 : ∀i = 1, · · · , N , z.ei < r0 .
Notons alors K̃0 l’ensemble
K̃0 = Ω ∩ Wr0 .
L’ensemble Ω" est donné par Ω" = "Ω. On peut remarquer que Ω" tend vers le réseau G quand " tend
vers 0. L’ensemble des trajectoires admissibles partant de z0 ∈ Ω" est donné par
Tz0 ," =
yz0 ∈ Li p(R+ ; Ω" ) : ∃α ∈ A t.q. yz0 (t) = x +
Z
t
α(s)ds ,
0
où A désigne l’ensemble des fonctions mesurables de R+ dans A un sous-ensemble compact de R2 . Le
coût instantané `" : Ω" → R est de la forme
¨
`" (z, a) =
`i (zi − "r0 ,
`0 ( "z , a)
zi⊥
" , a)
si z ∈ "S i ,
si z ∈ " K̃ 0 ,
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
47
où pour z ∈ R2 , (zi , zi⊥ ) sont les coordonnées de z dans la base {ei , ei⊥ }, et les fonctions `i = [0, +∞) ×
[−1, 1] × A → R (i = 1, · · · , N ) et `0 : K̃ 0 × A → R sont données. Les auteurs font l’hypothèse supplémentaire que la fonction `" ainsi définie est continue. Ils font également l’hypothèse de contrôlabilité :
il existe r > 0 tel que B(0, r) ⊂ A. La fonction valeur du problème de contrôle optimal est donnée par
v" (z0 ) = inf
z∈Tz0 ,"
Z
+∞
`" (z(t; z0 , α), α(t))e−λt d t,
0
et est solution de l’équation de Hamilton-Jacobi avec contraintes d’état
λu" (z) + H" (z, Du" (z)) ≥ 0
λu" (z) + H" (z, Du" (z)) ≤ 0
dans Ω" ,
dans Ω" ,
où le Hamiltonien H" : Ω" × R2 → R est défini par
H" (z, p) = max (−p.a − `" (z, a)) .
a∈A
Les auteurs prouvent que la suite de fonctions valeurs v" converge quand " tends vers 0 vers l’unique
solution de l’équation effective
λu(t, x) + H i (x, ddux i (x)) = 0
€
Š
du
λu(t, x) + max A, H O ( ddu
x 1 (x), · · · , d x N (x)) = 0
dans Ji \ {O},
sur {0},
(1.4.13)
où les Hamiltoniens effectifs H i et H O sont connus explicitement et où la constante A est un limiteur de
flux effectif. En particulier, le Hamiltonien H O peut s’écrire sous la forme
€ +
Š
H O (p1 , · · · , pN ) = max H i (pi ) ,
i=1,··· ,N
+
où le Hamiltonien H i correspond à la partie décroissante du Hamiltonien H i . Finalement, les Hamiltoniens effectifs H i et H O vérifient les bonnes propriétés pour que (1.4.13) entre dans le cadre d’application de [AOT15b].
L’équation effective en dehors du point de jonction O est obtenue en appliquant directement des résultats de O. Alvarez et M. Bardi dans [AB02]. La détermination de la condition de jonction est nouvelle. Y.
Achdou et N. Tchou doivent d’abord déterminer la constante ergodique A. Cette dernière est construite
ρ
comme limite d’une suite de constantes ergodiques A associées à des sous-domaines bornés Ωρ ⊂ Ω
avec coût instantané `0 . La difficulté majeure rencontrée est alors dans la construction de correcteurs à
la jonction. Ceci est dû au fait que le domaine Ω n’est pas borné. Pour remédier à ce problème, Y. Achdou et N. Tchou construisent des correcteurs sur des sous-domaines non-bornés de Ω qui pourront être
utilisés dans la méthode de la fonction-test perturbée de Evans. La stratégie développée pour construire
ces correcteurs ressemble à celle développée par H. Ishii dans [Ish13], avec à nouveau une complication
due au caractère non borné du domaine Ω. Finalement, la preuve du résultat de convergence utilise
la méthode de la fonction-test perturbée de Evans, avec quelques adaptations car les correcteurs à la
jonction sont non standard.
Dans [GIM14] G. Galise, C. Imbert et R. Monneau s’intéressent à un problème d’homogénéisation
précisé pour des équations de Hamilton-Jacobi définies sur R. Plus précisément, les auteurs considèrent
la famille d’équations de Hamilton-Jacobi
∂ t u" + H( "t , "x , u"x ) = 0
u" (O, x) = u0 (x)
sur (0, T ) × R,
sur × R,
(1.4.14)
où le Hamiltonien H : R3 → R est supposé satisfaire les hypothèses suivantes :
. Continuité : H ∈ C 0 (R3 ) ;
. Périodicité par rapport au temps : ∀k ∈ Z et ∀(t, x, p) ∈ R3 on a H(t + k, x, p) = H(t, x, p) ;
48
CHAPITRE 1. INTRODUCTION
. Régularité par rapport à la variable t : il existe un module de continuité ω tel que ∀s, t, x, p ∈ R
on a H(t, x, p) − H(s, x, p) ≤ ω(|t − s|(1 + max(H(s, x, p), 0))) ;
. Coercivité uniforme : lim|q|→+∞ H(t, x, q) = +∞ ;
. Quasi-convexité de H loin de l’origine : il existe ρ0 > 0 tel que ∀x ∈ R \ (−ρ0 , ρ0 ) il existe une
application continue t 7−→ p0 (t, x) tel que
H(t, x, ·) est décroissant sur (−∞, p0 (t, x)),
H(t, x, ·) est croissant sur (p0 (t, x), +∞);
. Profils de H loin de l’origine : il existe deux Hamiltoniens Hα (t, x, p), α = L, R, tels que
H(t, x + k, p) − H L (t, x, p) → 0 quand k ∈ Z tend vers − ∞,
H(t, x + k, p) − HR (t, x, p) → 0 quand k ∈ Z tend vers + ∞,
uniformément par rapport à (t, x, p) ∈ [0, 1]2 × R, et ∀k, j ∈ Z, (t, x, p) ∈ R3 et α = L, R on a
Hα (t + k, x + j, p) = Hα (t, x, p).
Si on note u" l’unique solution de (1.4.14), les auteurs souhaitent déterminer l’équation effective qui
caractérise la limite de la suite de fonction u" . L’intuition nous dit que cette équation effective devrait
impliquer des Hamiltoniens effectifs H α , α = L, R, de part et d’autre de 0 et une condition de jonction
en 0. En supposant que les Hamiltoniens effectifs ont d’assez bonnes propriétés, l’équation effective
pourrait être étudiée en utilisant les outils développés par C. Imbert et R. Monneau dans [IM14a].
Afin d’assurer de telles propriétés pour les Hamiltoniens effectifs H α , les auteurs supposent que les
Hamiltoniens Hα vérifient l’une des deux hypothèses ci-dessous.
. Quasi-convexité : pour tout α = L, R, Hα est indépendant du temps et il existe pα0 (indépendant
de (t, x)) tel que
Hα (x, ·) est décroissant sur (−∞, pα0 ),
Hα (x, ·) est croissant sur (pα0 , +∞);
. Convexité : pour tout α = L, R et pour tout (x, t) ∈ R2 , l’application p 7−→ Hα (t, x, p) est convexe.
Dans ce cadre, les auteurs parviennent à prouver que la suite de fonction u" converge vers l’unique
solution de l’équation effective
∂ t u(t, x) + H L (∂ x u(t, x)) = 0
∂ t u(t, x) + H R (∂ x u(t, x)) = 0
∂ t u(t, 0) + FA(∂ x u(t, 0)) = 0
u(t, O) = u0 (x)
pour
pour
pour
pour
t > 0, x < 0,
t > 0, x > 0,
t > 0, x = 0,
x ∈ R.
(1.4.15)
Pour bien comprendre l’équation (1.4.15) ci-dessus, il faut penser R comme étant un réseau constitué du sommet 0 et de deux arrêtes (−∞, 0] et [0, +∞). Alors l’équation (1.4.15) est à considérer comme étant analogue à l’équation (1.3.5) qui apparaît dans [IM14a]. En particulier en x = 0,
∂ x u(t, 0) = (∂ x u(t, 0− ), ∂ x u(t, 0+ )). La fonction de jonction effective FA : R2 → R est définie comme
dans (1.3.7) (en faisant attention aux conventions de signe) à partir des Hamiltoniens H L et H R et du
limiteur de flux effectif A. G. Galise et al démontrent que le limiteur de flux effectif A qui apparaît dans
la condition de jonction est toujours supérieur à A0 = maxα=L,R minq∈R H α (q). Ainsi, le limiteur de flux
A influx effectivement sur le comportement limite.
La difficulté essentielle pour déterminer l’équation effective (1.4.15) réside dans la détermination de
la condition de jonction, et plus précisément du limiteur de flux effectif A. Comme dans [AT15], les
auteurs prouvent que le limiteur de flux A peut être obtenu comme la limite d’une suite de constantes
ρ
ergodique A associées à des problèmes de cellule définis sur des domaines tronqués R × [−ρ, ρ] ⊂ R2 .
La nouveauté apportée ici par G. Galise et al est la construction un correcteur global au point de jonction,
qui pourra être utilisé dans la méthode de la fonction-test perturbée de Evans. En effet, ils prouvent
par des arguments de stabilité que les correcteurs wρ associés aux problèmes de cellules tronquées
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
49
convergent vers un correcteur global défini sur R2 . Plus précisément, si on note w la limite des correcteurs tronqués wρ , les auteurs prouvent que le couple (w, A) est solution du problème de cellule défini
sur R2 . On remarque que dans la preuve de la convergence, les auteurs utilisent un résultat technique
prouvé par C. Imbert et al dans [IM14a] qui leur permet de réduire considérablement l’espace des
fonctions-test pour lequel la condition de jonction doit être vérifiée.
Dans [FS14a] N. Forcadel et W. Salazar souhaitent déterminer le modèle de trafic routier macroscopique qui découle d’un modèle microscopique contenant une perturbation qui ralentit certains véhicules
dans une région localisée de l’espace. L’objectif est de comprendre l’influence d’une perturbation locale
sur le comportement macroscopique du trafic. Le modèle microscopique est un modèle dans lequel les
véhicules se meuvent sur une ligne droite et où l’évolution du j-ème véhicule est régie par l’équation
différentielle ordinaire du premier ordre
U j0 (t) = V (U j+1 (t) − U j (t)).φ(U j (t)).
Dans cette équation U j désigne la position du j-ème véhicule, U j0 sa vitesse et où V (.) est la fonction
de vitesse optimale. La fonction φ : R → [0, 1] permet de modéliser la présence d’une perturbation localisée dans un voisinage de l’origine. Afin de déterminer le modèle macroscopique associé au modèle
microscopique considéré ici, les auteurs introduisent la fonction de répartition des véhicules redimensionnée ρ " (s’exprimant en fonction des Ui et de ") qui peut être caractérisée comme étant l’unique
solution d’une équation aux dérivées partielles de la forme
” " —
u (t,.)
(x).φ "x .|∂ x u" (t, x)| = 0 sur (0, +∞) × R,
∂ t u" (t, x) + M "
"
(1.4.16)
u" (0, x) = u0 (x)
sur R,
où M " [.] est un opérateur non-local qui s’exprime à l’aide de la fonction de vitesse optimale V (.).
Les auteurs souhaitent déterminer l’équation effective qui caractérise la limite de la suite de fonctions
(ρ " ). Comme la perturbation est localisée dans un voisinage de l’origine, on s’attend à trouver une
équation effective de type équation de Hamilton-Jacobi avec une condition de jonction à l’origine faisant
intervenir un limiteur de flux effectif. Malgré la présence du terme non-local dans (1.4.16), les auteurs
parviennent à adapter les méthodes développées dans [AT15] et [GIM14] pour étudier le comportement
asymptotique des fonctions ρ " quand " tend vers 0. En particulier, la détermination du limiteur de flux
effectif ainsi que la construction d’un correcteur global à l’origine font intervenir des problèmes de
cellule tronqués. Ainsi N. Forcadel et W. Salazar parviennent à prouver que la suite de fonctions ρ "
converge quand " tend vers 0 vers l’unique solution u0 d’une équation de Hamilton-Jacobi effective de
la forme
∂ t u0 (t, x) + H(∂ x u0 (t, x)) = 0
∂ t u0 (t, x) + H(∂ x u0 (t, x)) = 0
∂ t u0 (t, x) + FA(∂ x u0 (t, 0− ), ∂ x u0 (t, 0+ )) = 0
u0 (0, x) = u0 (x)
sur
sur
sur
sur
(0, +∞) × (−∞, 0),
(0, +∞) × (0, +∞),
(0, +∞) × {0},
R.
(1.4.17)
où FA est une fonction de jonction effective définie de manière analogue à (1.3.7) avec un limiteur de
flux effectif A.
1.4.3
Résumé du Chapitre 4
Ce paragraphe est consacré à la présentation des travaux qui apparaissent dans la Partie III de ce manuscrit, qui s’inscrit dans le développement de la théorie de l’homogénéisation des équations de HamiltonJacobi discontinues.
Le Chapitre 4 est un travail en commun avec Y. Achdou et N. Tchou qui fait l’objet d’une prépublication [AOT15a]. Nous nous intéressons à une famille de problèmes de contrôle optimal à horizon
infini dans R2 pour lesquels les dynamiques et les coûts instantanés peuvent être discontinus à travers
une frontière oscillante Γ" . On suppose que les oscillations des frontières oscillantes Γ" sont de période
et d’amplitude du même ordre de grandeur ". Grâce à l’étude effectuée dans [Oud14], nous sommes
en mesure de caractériser les fonctions valeur v" par des équations de Hamilton-Jacobi-Bellman. On
50
CHAPITRE 1. INTRODUCTION
souhaite alors déterminer l’équation effective qui caractérise la limite des fonctions valeur v" quand "
tend vers 0.
Commençons par introduire plus précisément le cadre. La frontière oscillante Γ" est définie par
n
x o
2
,
Γ" = (x 1 , x 2 ) ∈ R2 : x 1 = " g
"
où g : R → R est une fonction 1-périodique de classe C 2 . On note Ω"L le demi-espace situé à gauche de
Γ" et ΩR" celui situé à sa droite.
n
x o
2
Ω"L = (x 1 , x 2 ) ∈ R2 : x 1 < " g
"
et
n
x o
2
ΩR" = (x 1 , x 2 ) ∈ R2 : x 1 > " g
.
"
On considère les dynamiques f i : R2 × Ai → R2 et les coûts instantanés `i : R2 × Ai → R associés
à chaque demi-espace Ω"i , où les Ai sont des sous-espaces compact d’un espace métrique A. On peut
toujours supposer que les ensembles Ai sont disjoints. Les hypothèses prises sur les fonctions f i et
`i sont semblables à celles qui ont été faites dans [AOT15b] et [Oud14]. On précise tout de même
l’hypothèse de contrôlabilité qui joue un rôle très important dans notre étude : il existe δ0 > 0 tel que
pour tout i = L, R et x ∈ Γ" on a
B(0, δ0 ) ⊂ f i (x, a) : A ∈ Ai .
Pour définir la dynamique et le coût instantané sur R2 entier, nous introduisons l’ensemble
M" = (x, a); x ∈ R2 , a ∈ Ai si x ∈ Ω"i , i = L, R, et a ∈ AL ∪ AR si x ∈ Γ" .
On définit alors la dynamique f" sur M" par
∀(x, a) ∈ M" ,
f" (x, a) =
§
f i (x, a)
f i (x, a)
si x ∈ Ω"i \Γ" ,
si x ∈ Γ" et a ∈ Ai ,
`" (x, a) =
§
`i (x, a)
`i (x, a)
si x ∈ Ω"i \Γ" ,
si x ∈ Γ" et a ∈ Ai .
et le coût instantané `" par
∀(x, a) ∈ M" ,
Les fonctions f" et `" sont bien définies et continues sur M" puisqu’on a supposé que les ensembles
Ai sont disjoints. L’ensemble des trajectoires admissibles partant de x de notre problème de contrôle
optimal est donné par


∞
+
y x ∈ Lip(R+ ; R2 ),
 ( y x , a) ∈ Lloc (R ; M" ) :

Z t
T x," =
,
y x (t) = x +
f" ( y x (s), a(s))ds ∀t ∈ R+ 

0
et la fonction valeur s’écrit
v" (x) =
Z
∞
inf
( y x ,α)∈T x,"
`" ( y x (t), α(t))e−λt d t.
0
D’après [Oud14], on sait que la fonction valeur v" est une fonction continue et bornée et peut être caractérisée comme solution de viscosité d’une équation de Hamilton-Jacobi-Bellman, avec une condition
de transmission sur Γ"
L’espace des fonctions-test est défini par
Définition 1.4.1. Une fonction φ : R2 → R est une (")-fonction-test admissible si
. φ est continue sur R2
. pour tout j = L, R, φ|
j
j
Ω"
∈ C 1 (Ω" ).
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
51
On note R" l’ensemble des (")-fonctions-test admissibles. Pour φ ∈ R" , x ∈ Γ" et i ∈ {L, R}, on notera
Dφ i (x) = lim x 0 →x Dφ(x 0 ).
x 0 ∈Ω"i
L’équation de Hamilton-Jacobi-Bellman qui caractérise la fonction valeur v" est alors donnée par
λu(x) + H i (x, Dφ i (x))
λu(x) + HΓ" (x, Dφ L (x), Dφ R (x))
= 0,
= 0,
si x ∈ Ω"i ,
si x ∈ Γ" ,
(1.4.18)
où pour i = L, R, les Hamiltoniens H i : R2 × R2 → R et HΓ" : Γ" × R2 × R2 → R sont définis par
H i (x, p)
=
max(−p · f i (x, a) − `i (x, a)),
HΓ" (x, p L , pR )
=
max{ HΓ" (x, p L ), HΓ" (x, pR )},
(1.4.19)
a∈Ai
+,L
+,R
(1.4.20)
où, en désignant par n"i (x) le vecteur unitaire normal à Γ" en x qui pointe vers Ω"i ,
+,i
HΓ" (x, p) =
a∈Ai
t.q.
max
(−p · f i (x, a) − `i (x, a)),
f i (x,a)·n"i (x)≥0
∀x ∈ Γ" , ∀p ∈ R2 .
Les sous- et sur-solutions de viscosité pour (1.4.18) sont définies comme dans la Définition 1.3.7, en
utilisant l’espace des fonctions-test R" , le Hamiltonien H i si x ∈ Ω"i et le Hamiltonien HΓ" si x ∈ Γ" .
Travailler directement avec (1.4.18) est compliqué puisque la géométrie dépend du paramètre ". Il
sera plus pratique de redresser la géométrie et de travailler dans une géométrie fixe via le changement
de variable

x ‹
x − g( "2
,
z = G(x) = 1
x2
de Jacobienne

1
J" (x) =
0
−g 0 (
1
x2 ‹
" ) .
Ce changement de variable conduit à un nouveau problème de contrôle optimal dans R2 , faisant intervenir la frontière fixe
Γ = (x 1 , x 2 ) ∈ R2 : x 1 = 0 ,
et les demi-espaces
Ω L = (x 1 , x 2 ) ∈ R2 : x 1 < 0
et
ΩR = (x 1 , x 2 ) ∈ R2 : x 1 > 0 .
Le problème redressé implique une modification des dynamiques et des coûts instantanés de part et
d’autre de la frontière Γ : ceux-ci sont donnés par
f˜"i :
Ω̄i × Ai
(z, a)
→
7
→
R2
J" (z) f i G −1 (z), a ,
et
˜`i :
"
Ω̄i × Ai
(z, a)
→
7
→
R
`i G −1 (z), a .
Les fonctions f˜"i et ˜`"i vérifient des hypothèses semblables aux fonction f i et `i , dans la géométrie
redressée. Les analogues de M" , f" et `" dans cette nouvelle géométrie sont donnés par
M = (x, a); x ∈ R2 , a ∈ Ai si x ∈ Ωi , et a ∈ AL ∪ AR si x ∈ Γ ,
f˜"i (z, a)
si x ∈ Ωi ,
∀(z, a) ∈ M ,
f˜" (z, a) =
i
˜
si x ∈ Γ et a ∈ Ai ,
f" (z, a)
i
˜
si x ∈ Ωi ,
˜`" (z, a) = `" (z, a)
∀(z, a) ∈ M ,
˜`i (z, a)
si x ∈ Γ et a ∈ Ai .
"
De même, l’ensemble des trajectoires admissibles partant de x ∈ R2 dans la nouvelle géométrie est
donné par


∞
+
y x ∈ Lip(R+ ; R2 ),
 ( y x , a) ∈ Lloc (R ; M ) :

Z t
Tex," =
+
y x (t) = x +
f˜" ( y x (s), a(s))ds ∀t ∈ R 

0
52
CHAPITRE 1. INTRODUCTION
et le nouveau problème de contrôle optimal consiste à trouver la fonction valeur
Z∞
˜`" ( yz (t), a(t))e−λt d t.
ṽ" (z) = inf
( yz ,a)∈Tez,"
0
2
On peut remarquer que pour tout z ∈ R , on a la relation
z ṽ" (z) = v" (G −1 (z)) = v" z1 + " g 2 , z2 .
"
(1.4.21)
L’espace des fonctions-test, les Hamiltoniens, les notions de sous- et sur-solution, sont transformés de
manière cohérentes avec le changement de variable.
Notre objectif est de comprendre le comportement asymptotique de la suite de fonctions valeur
(v" ) quand " tend vers 0. D’après la relation (1.4.21), la suite (v" ) converge si et seulement si la suite
(ṽ" ) converge. De plus, lorsqu’il y a convergence, les deux suites ont la même limite. Comme signalé
ci-dessus, il est plus pratique de se concentrer sur l’étude du comportement asymptotique de la suite
(ṽ" ), qui correspond à une suite de problèmes de contrôle pour lesquels la géométrie est fixée.
En travaillant dans la géométrie fixe, on arrive à prouver le résultat de convergence suivant :
Théorème 1.4.1 (Convergence). Quand " tend vers 0, v" (resp. ṽ" ) converge uniformément vers v l’unique
solution de viscosité de l’équation effective
λv(z) + H i (z, Dv(z)) = 0
λv(z) + max A(z2 , ∂z2 v(z)), HΓ (z, Dv L (z), Dv R (z)) = 0
si z ∈ Ωi ,
(1.4.22)
si z = (0, z2 ) ∈ Γ ,
(1.4.23)
où le Hamiltonien H i est défini par (1.4.19), le Hamiltonien HΓ : Γ × R2 × R2 → R est défini par
HΓ (z, p L , pR ) = max H +,L (z, p L ), H +,R (z, pR ) ,
avec, en désignant par ni le vecteur unitaire normal à Γ en x qui pointe vers Ωi ,
H +,i (z, p) =
a∈Ai
max
(−p. f i (z, a) − `i (z, a)),
t.q. f i (z,a).ni ≥0
et A : R2 → R est un Hamiltonien tangentiel effectif.
Avant d’expliquer la preuve du résultat de convergence, détaillons la construction du limiteur de
flux A effectif. Comme dans [AT15] et [GIM14], on est confronté à des difficultés pour construire des
correcteurs globaux dans R2 qui est non-borné. On suit donc la démarche présentée dans [AT15] qui
consiste à d’abord considérer des problèmes de cellule définis sur des domaines tronqués [−ρ, ρ] × R.
Plus précisément, pour z = (0, z2 ) ∈ Γ , p2 ∈ R et ρ > 0 fixé, on considère le problème de cellule tronquée
ρ
 L
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 )




H̃ R ((0, z2 ), Du( y) + p2 e2 , y2 )


 max H̃ +,i ((0, z ), Dui ( y) + p e , y )
2
2 2
2
= A (z2 , p2 )
ρ
= A (z2 , p2 )
ρ
= A (z2 , p2 )

H̃ −,L ((0, z2 ), Du( y) + p2 e2 , y2 )



−,R


 H̃ ((0, z2 ), Du( y) + p2 e2 , y2 )
u est 1-périodique par rapport à y2 , .
= A (z2 , p2 )
ρ
= A (z2 , p2 )
i=L,R
où si J̃( y2 ) =

1
0
ρ
dans (−ρ, 0) × R,
dans (0, ρ) × R,
sur Γ ,
(1.4.24)
sur {−ρ} × R,
sur {ρ} × R,
‹
−g 0 ( y2 )
le Hamiltonien H̃ i : R2 × R2 × R → R est défini par
1
H̃ i (z, p, y2 ) := max −J̃( y2 ) f i (z, a) · p − `i (z, a) ,
a∈Ai
et si ni désigne le vecteur unitaire normal à Γ en x qui pointe vers Ωi , les Hamiltoniens H̃ +,i et H̃ −,i :
R2 × R2 × R → R sont définis par
H̃ ±,i (z, p, y2 ) =
max
−J̃( y2 ) f i (z, a) · p − `i (z, a) .
a ∈ Ai t.q.
±J̃( y2 ) f i (z, a) · ni ≥ 0
1.4. HOMOGÉNÉISATION ET PERTURBATIONS SINGULIÈRES
53
Les conditions qui apparaissent sur le bord droit du domaine {ρ}×R dans l’équation (1.4.24) signifient
qu’on sélectionne uniquement les contrôles qui permettent de rester dans le domaine [−ρ, ρ] × R. On
peut faire le même commentaire au sujet de la condition qui apparait sur le bord gauche du domaine
{−ρ} × R dans l’équation (1.4.24).
Proposition 1.4.1. Pour tout z = (0, z2 ) ∈ Γ , p2 ∈ R et ρ > 0 fixés, il existe une unique constante
ρ
A (z2 , p2 ) ∈ R telle que l’équation (1.4.24) admet une solution wρ (z2 , p2 , ·).
La preuve de ce résultat utilise fortement les résultats d’existence et le principe de comparaison
ρ
prouvés dans [Oud14]. On peut facilement montrer que la suite A (z2 , p2 ) est croissante et bornée et
donc convergente. En conséquence :
Définition 1.4.2. On définit le Hamiltonien tangentiel effectif A(z2 , p2 ) par
ρ
A(z2 , p2 ) = lim A (z2 , p2 ).
ρ→∞
On prouve alors que le Hamiltonien tangentiel effectif A satisfait de bonnes propriétés (continuité,
convexité et coercivité), ce qui permet, en application des résultats de [Oud14], de déduire que l’équation (1.4.22)-(1.4.23) admet au plus une solution. On prouve également que pour tout z Z , p2 ∈ R, on
a
A(z2 , p2 ) ≥ A0 (z2 , p2 ) = max inf H i ((0, z2 ), p2 e2 + pe1 ).
i=L,R q∈R
Ceci nous assure que le Hamiltonien tangentiel effectif A est bien actif dans l’équation (1.4.23).
On utilise maintenant l’idée proposée par G. Galise et al dans [GIM14] pour construire des correcteurs
globaux qui pourront servir dans la méthode de la fonction test-perturbée de Evans, [Eva89]. En effet,
ρ
si (wρ (z2 , p2 , ·), A (z2 , p2 )) est une solution du problème de cellule tronquée (1.4.24), on montre par
des arguments de stabilité la convergence vers une solution du problème de cellule global
 i
 H̃ ((0, z2 ), Du( y) + p2 e2 , y2 )
max H̃ +,i ((0, z2 ), Dui ( y) + p2 e2 , y2 )
 i=L,R
u est 1-périodique par rapport à y2 , .
= A(z2 , p2 )
= A(z2 , p2 )
dans Ωi ,
sur Γ ,
(1.4.25)
Proposition 1.4.2. Pour tout z = (0, z2 ) ∈ Γ , p2 ∈ R fixés, l’équation (1.4.25) admet une solution
w(z2 , p2 , ·).
Nous avons maintenant assez d’éléments pour expliquer la preuve du résultat de convergence. En
dehors des points d’interface, cette preuve est standard. Pour justifier la condition de transmission
(1.4.23), on applique la méthode de la fonction-test perturbée de Evans. On utilise deux résultats importants pour conclure lors de l’application de cette méthode. D’une part, on applique le résultat prouvé
par C. Imbert et al dans [IM14a] qui permet de réduire considérablement l’espace des fonctions-test
pour la condition de jonction (1.4.23). D’autre part, on prouve des propriétés qualitatives sur le comportement asymptotique de la fonction "w(z2 , p2 , "· ) quand " tend vers 0. Plus précisément, on montre
que la suite de fonctions "w(z2 , p2 , "· ) tend vers une fonction W (z2 , p2 , ·) quand " tend vers 0, qui est
indépendante de y2 et pour laquelle on dispose de contrôles sur les pentes (c’est-à-dire ∂ y1 W (z2 , p2 , ·)).
54
1.5
CHAPITRE 1. INTRODUCTION
Organisation du manuscrit
Ce manuscrit est divisé en deux parties selon les deux axes de cette thèse : l’étude d’équations de
Hamilton-Jacobi-Bellman discontinues sur des structures de type réseau, puis celle de problèmes de
perturbation singulière impliquant des équations de Hamilton-Jacobi discontinues.
Chapitre 2 Dans ce chapitre nous nous intéressons à des problèmes de contrôle optimal à horizon
infini sur des réseaux, pour lesquels on autorise différentes dynamiques et différents coûts instantanés
sur chaque arête du réseau. Comme dans les cas plus classiques, nous montrons que la fonction valeur
d’un tel problème peut être caractérisée par le biais d’une équation de Hamilton-Jacobi-Bellman définie
sur le réseau, et qui fait intervenir des conditions de jonction aux sommets du réseau.
Chapitre 4 Dans ce chapitre nous généralisons les résultats obtenus dans le Chapitre 2 au cas
de problèmes de contrôle optimal à horizon infini sur une muti-jonction, pour lesquels on autorise
différentes dynamiques et différents coûts instantanés sur chaque branche de la multi-jonction. Nous
traitons ce problème d’abord sous une hypothèse de contrôlabilité forte, puis sous une hypothèse de
contrôlabilité plus faible, dite hypothèse de contrôlabilité normale.
Chapitre 3 Dans ce chapitre nous étudions une famille de problèmes de contrôle optimal à horizon
infini dans le plan où la dynamique et le coût instantané peuvent être discontinus à travers une frontière
oscillante Γ" . Nous montrons que la suite des fonctions valeurs v" converge quand " tend vers 0 vers
l’unique solution v d’une équation de Hamilton-Jacobi-Bellman effective avec une condition de jonction
effective qui conserve une mémoire des oscillations de la frontière Γ" .
DEUXIÈME
PARTIE
Hamilton-Jacobi-Bellman equations
on heterogeneous structures with
geometric singularity
55
56
CHAPITRE 2
HAMILTON-JACOBI EQUATIONS FOR
OPTIMAL CONTROL ON JUNCTIONS AND
NETWORKS
Résumé
On considère des problèmes de contrôle optimal pour lesquels l’état est
contraint à rester sur un réseau. Une notion de solution de viscosité des
équations de Hamilton-Jacobi associées a été proposée dans des travaux antérieurs. Ici, on propose une preuve simple d’un principe de comparaison.
Cette preuve est basée sur des arguments de contrôle optimal. La stabilité
des solutions de viscosité est aussi étudiée.
Abstract
We consider continuous-state and continuous-time control problems where
the admissible trajectories of the system are constrained to remain on a network. A notion of viscosity solution of Hamilton-Jacobi equations on the
network has been proposed in earlier articles. Here, we propose a simple
proof of a comparison principle based on arguments from the theory of optimal control. We also discuss stability of viscosity solutions.
Introduction
A network (or a graph) is a set of items, referred to as vertices (or nodes/crosspoints), with connections
between them referred to as edges. In the recent years there has been an increasing interest in the
investigation of dynamical system and differential equation on networks, in particular in connection
with problem of data transmission and traffic management (see for example Garavello-Piccoli [GP06],
Engel et al. [EKFNS08]). While control problems with state constrained in closures of open sets are
well studied ( [Son86a, Son86b], [CDL90], [IK96]) there is to our knowledge much fewer literature
on problems on networks. The results of Frankowska and Plaskacz [FP00b, FP00a] do apply to some
closed sets with empty interior, but not to networks with crosspoints (except in very particular cases).
The literature on continuous-state and continuous-time control on networks is recent: the first two
articles were published in 2012: control problems whose dynamics is constrained to a network and related Hamilton-Jacobi equations were studied in [ACCT13]: a Hamilton-Jacobi equation on the network
was proposed, with a definition of viscosity solution, which reduces to the usual one if the network is a
straight line (i.e. is composed of two parallel edges sharing an endpoint) and if the dynamics and cost
are continuous; while in the interior of an edge, one can test the equation with a smooth test-function,
the main difficulties arise at the vertices where the network does not have a regular differential structure. At a vertex, a notion of derivative similar to that of Dini’s derivative (see for example [BCD97])
was proposed: admissible test-functions are continuous functions whose restriction to each edge is C 1 .
With this definition, the intrinsic geodesic distance, fixed one argument, is an admissible test-function
with respect to the other argument. The Hamiltonian at a vertex depends on all directional derivatives
in the directions of the edges containing the vertex, see § 2.2.3 below. Independently, Imbert, Monneau
and Zidani [IMZ13] proposed an equivalent notion of viscosity solution for studying a Hamilton-Jacobi
approach to junction problems and traffic flows. There is also the work by Schieborn and Camilli [SC13],
in which the authors focus on eikonal equations on networks and on a less general notion of viscosity
solution.
57
58
CHAPITRE 2. THE NETWORK CASE
Both [ACCT13] and [IMZ13] contain the first comparison and uniqueness results: in [ACCT13], suitably modified geodesic distances are used in the doubling variables method for proving comparison
theorems under rather strong continuity assumptions. In [IMZ13], Imbert, Monneau and Zidani used
a completely different argument based on the explicit solution of a related optimal control problem,
which could be obtained because it was assumed that the Hamiltonians associated with each edge did
not depend on the state variable.
A general comparison result has finally been obtained in the quite recent paper by Imbert-Monneau
[IM14a]. In the latter article, the Hamiltonians in the edges are completely independent from each
other; the main assumption is that the Hamiltonian in each edge, say H i (x, p) for the edge indexed
i, is bimonotone, i.e. non increasing (resp. non decreasing) for p smaller (resp. larger) than a given
threshold pi0 (x). Of course, convex Hamiltonian coming from optimal control theory are bimonotone.
Moreover, [IM14a] handles more general transmission conditions than the previous articles, with an
additional running cost at the junctions. In [IM14a], the proof of the comparison result is rather involved and only uses arguments from the theory of partial differential equations: in the most simple
case where all the Hamiltonians related to the edges are strictly convex and reach their minima at p = 0,
the idea consists of doubling the variables and using a suitable test-function; then, in the general case,
perturbation arguments are used for applying the results proved in the former case.
In coincidence with these research efforts about networks, Barles, Briani and Chasseigne, see [BBC13,
BBC14], have recently studied control problems with discontinuous dynamics and costs, obtaining comparison results for some Bellman equations arising in this context, with original and elegant arguments.
Related problems were also recently addressed by Rao, Siconolfi and Zidani [RZ13, RSZ14].
The aim of the present paper is to focus on optimal control problems with independent dynamics and
running costs in the edges, and to show that the arguments in [BBC13] can be adapted to yield a simple
proof of a comparison result.
Sections 2.1 to 2.4 are devoted to the case of a junction, i.e. a network with one vertex only. Section 2.1 contains a description of the geometry and of the optimal control problem. In Section 2.2, a
Hamilton-Jacobi equation is proposed for the value function, together with a notion of viscosity solution. It is proved that the value function is a viscosity solution of the Hamilton-Jacobi equation. Also in
Section 2.2, Lemma 2.2.1 on the structure of the Hamiltonian at the vertex will be important for obtaining the comparison principle. Some important properties of viscosity sub and supersolutions are given
in Section 2.3, and the comparison principle is proved in Section 2.4. In §2.5 we discuss the stability of
the viscosity sub and super solution under perturbations of the Hamiltonians. In §2.6 we show that all
the results can be easily extended to the case when there is an additional cost at the junction. Finally,
in Section 2.7, the results obtained for the junction are generalized for networks with more than one
vertices.
2.1
The junction
2.1.1
The geometry
Let us focus on the model case of a junction in Rd with N semi-infinite straight edges, N > 1. The
edges are denoted by (Ji )i=1,...,N . The edge Ji is the closed half-line R+ ei . The vectors ei are two by two
distinct unit vectors in Rd . The half-lines Ji are glued at the origin O to form the junction G :
G=
N
[
Ji .
i=1
The geodetic distance d(x, y) between two points x, y of G is
§
|x − y| if x, y belong to the same edge Ji
d(x, y) =
|x| + | y| if x, y belong to different branches Ji and J j .
2.1.2
The optimal control problem
We consider infinite horizon optimal control problems which have different dynamics and running costs
in the edges. We are going to describe the assumptions on the dynamics and costs in each edge Ji . The
2.1. THE JUNCTION
59
sets of controls are denoted by Ai and the system is driven by a dynamics f i and the running cost is
given by `i . Our main assumptions are as follows
[H0] A is a metric space (one can take A = Rm ). For i = 1, . . . , N , Ai is a non empty compact subset of
A and f i : Ji × Ai → R is a continuous bounded function. The sets Ai are disjoint. Moreover, there
exists L > 0 such that for any i, x, y in Ji and a ∈ Ai ,
| f i (x, a) − f i ( y, a)| ≤ L|x − y|.
We will use the notation Fi (x) for the set { f i (x, a)ei , a ∈ Ai }.
[H1] For i = 1, . . . , N , the function `i : Ji × Ai → R is a continuous and bounded function. There is
a modulus of continuity ωi such that for all x, y in Ji and for all a ∈ Ai , |`i (x, a) − `i ( y, a)| ≤
ωi (|x − y|).
[H2] For i = 1, . . . , N , x ∈ Ji , the non empty and closed set
FLi (x) ≡ {( f i (x, a)ei , `i (x, a)), a ∈ Ai }
is convex.
[H3] There is a real number δ > 0 such that for any i = 1, . . . , N ,
[−δei , δei ] ⊂ Fi (O).
Remark 2.1.1. In [H0] the assumption that the sets Ai are disjoint is not restrictive: it is made only for
simplifying the proof of Theorem 2.1.2 below. Indeed, if Ai are not disjoint, then we define Ãi = {i} × Ai
and ˜f i (x, ã) = f i (x, a), `˜i (x, ã) = `i (x, a) if x ∈ Ji and ã = (i, a) with a ∈ Ai .
The sets Ãi are disjoint compact subsets of à = ∪i=1,...,n Ãi which is a (compact) metric space for the distance
d̃((i, a), ( j, b)) = |i − j| + dA(a, b), and the functions ˜f i , `˜i inherit the properties of f i and `i .
The assumption [H2] is not essential: it is made in order to avoid the use of relaxed controls.
Assumption [H3] is a strong controllability condition at the vertex (which implies the coercivity of the
Hamiltonian). It has already been widely used in the framework of networks (for instance, the same
assumption is made in [ACCT13, BBC13], and the coercivity of the Hamiltonian is assumed [IM14a]).
We will see that [H3] yields the continuity of the value function. Without any controllability condition,
the value function may not be continuous and the definition of the viscosity solutions should differ
from the one proposed below. There are of course milder controllability conditions, but with them, our
techniques do not seem to apply in a straightforward manner.
Here is a general version of Filippov implicit function lemma, see [MW67], which will be useful to prove
Theorem 2.1.2 below.
Theorem 2.1.1. Let I be an interval of R and γ : I → Rd × Rd be a measurable function. Let K be a closed
subset of Rd × A and Ψ : K → Rd × Rd be continuous. Assume that γ(I) ⊂ Ψ(K), then there is a measurable
function Φ : I → K with
Ψ ◦ Φ(t) = γ(t) for a.a. t ∈ I.
Proof. See [MW67].
Let us denote by M the set:
M = (x, a); x ∈ G ,
a ∈ Ai if x ∈ Ji \{O}, and a ∈ ∪Ni=1 Ai if x = O .
The set M is closed. We also define the function f on M by
§
f i (x, a)ei
∀(x, a) ∈ M ,
f (x, a) =
f i (O, a)ei
if x ∈ Ji \{O},
if x = O and a ∈ Ai .
The function f is continuous on M because the sets Ai are disjoint. Let F̃ (x) be defined by
§
Fi (x)
if x belongs to the edge Ji \{O}
F̃ (x) =
∪Ni=1 Fi (O)
if x = O.
(2.1.1)
60
CHAPITRE 2. THE NETWORK CASE
For x ∈ G , the set of admissible trajectories starting from x is
Yx =
ẏ (t) ∈ F̃ ( y (t)),
x
y x ∈ Li p(R ; G ) : x
y x (0) = x,
+
for a.a. t > 0,
.
(2.1.2)
Theorem 2.1.2. Assume [H0],[H1],[H2] and [H3]. Then
1. For any x ∈ G , Yx is non empty.
2. For any x ∈ G , for each trajectory y x in Yx , there exists a measurable function Φ : [0, +∞) → M ,
Φ(t) = (φ1 (t), φ2 (t)) with
( y x (t), ẏ x (t)) = (φ1 (t), f (φ1 (t), φ2 (t))),
for a.e. t,
which means in particular that y x is a continuous representation of φ1
3. Almost everywhere in [0, +∞),
ẏ x (t) =
N
X
1{ y x (t)∈Ji \{O}} f i ( y x (t), φ2 (t))ei .
i=1
4. Almost everywhere on {t : y x (t) = O}, f (O, φ2 (t)) = 0.
Proof. The proof of point 1 is easy, because 0 ∈ F̃ (O).
The proof of point 2 is a consequence of Theorem 2.1.1, with K = M , I = [0, +∞), γ(t) = ( y x (t), ẏ x (t))
and Ψ(x, a) = (x, f (x, a)).
From point 2, we deduce that
ẏ x (t) =
N
X
1{ y x (t)∈Ji \{O}} f i ( y x (t), φ2 (t))ei + 1{ y x (t)=O}} f (O, φ2 (t)),
i=1
and from Stampacchia’s theorem, f (O, φ2 (t)) = 0 almost everywhere in {t : y x (t) = O}. This yields
points 3 and 4.
It is worth noticing that in Theorem 2.1.2, a solution y x can be associated with several control laws
φ2 (·). We introduce the set of admissible controlled trajectories starting from the initial datum x:
Tx =

∞
+
 ( y x , α) ∈ LLoc (R ; M ) :

y x ∈ Lip(R+ ; G ),
Z t
y x (t) = x +
f ( y x (s), α(s))ds
0


in R+ 
.
(2.1.3)
Remark 2.1.2. If two different edges are aligned with each other, say the edges J1 and J2 , many other
assumptions can be made on the dynamics and costs:
• a trivial case in which the assumptions [H1]-[H3] are satisfied is when the dynamics and costs are
continuous at the origin, i.e. A1 = A2 ; f1 and f2 are respectively the restrictions to J1 ×A1 and J2 ×A2
of a continuous and bounded function f1,2 defined in Re1 × A1 , which is Lipschitz continuous with
respect to the first variable; `1 and `2 are respectively the restrictions to J1 × A1 and J2 × A2 of a
continuous and bounded function `1,2 defined in Re1 × A1 .
• In this particular geometrical setting, one can allow some mixing (relaxation) at the vertex with
several possible rules: More precisely, in [BBC13, BBC14], Barles et al. introduce several kinds of
trajectories which stay at the junction: the regular trajectories are obtained by mixing outgoing dynamics from J1 and J2 , whereas singular trajectories are obtained by mixing strictly ingoing dynamics
from J1 and J2 . Two different value functions are obtained whether singular mixing is permitted or
not.
2.2. THE HAMILTON-JACOBI EQUATION
The cost functional
61
The cost associated to the trajectory ( y x , α) ∈ T x is
J(x; ( y x , α)) =
Z
∞
`( y x (t), α(t))e−λt d t,
(2.1.4)
0
where λ > 0 is a real number and the Lagrangian ` is defined on M by
∀(x, a) ∈ M ,
The value function
`(x, a) =
§
`i (x, a)
`i (O, a)
if x ∈ Ji \{O},
if x = O and a ∈ Ai .
The value function of the infinite horizon optimal control problem is
v(x) =
inf
( y x ,α)∈T x
J(x; ( y x , α)).
(2.1.5)
Proposition 2.1.1. Assume [H0],[H1],[H2] and [H3]. Then the value function v is bounded and continuous on G .
Proof. The proof essentially uses Assumption [H3]. Since it is classical, we skip it.
2.2
2.2.1
The Hamilton-Jacobi equation
Test-functions
For the definition of viscosity solutions on the irregular set G , it is necessary to first define a class of the
admissible test-functions
Definition 2.2.1. A function φ : G → R is an admissible test-function if
• φ is continuous in G and C 1 in G \ {O}
• for any j, j = 1, . . . , N , φ|J j ∈ C 1 (J j ).
The set of admissible test-functions is noted R(G ). If φ ∈ R(G ) and ζ ∈ R, let Dφ(x, ζei ) be defined by
dφ
dφ
Dφ(x, ζei ) = ζ d x i (x) if x ∈ Ji \{O} and Dφ(O, ζei ) = ζ limh→0+ d x i (hei ).
Property 2.2.1. If φ = g ◦ ψ with g ∈ C 1 and ψ ∈ R(G ), then φ ∈ R(G ) and
Dφ(O, ζ) = g 0 (ψ(O))Dψ(O, ζ).
2.2.2
Vector fields
For i = 1, . . . , N , we denote by Fi+ (O) and FL+i (O) the sets
Fi+ (O) = Fi (O) ∩ R+ ei ,
FL+i (O) = FLi (O) ∩ (R+ ei × R),
which are non empty thanks to assumption [H3]. Note that 0 ∈ ∩Ni=1 Fi (O). From assumption [H2],
these sets are compact and convex. For x ∈ G , the sets F (x) and FL(x) are defined by
F (x) =
§
Fi (x)
∪i=1,...,N Fi+ (O)
FL(x) =
§
FLi (x)
∪i=1,...,N FL+i (O)
if x belongs to the edge Ji \{O}
if x = O,
and
if x belongs to the edge Ji \{O}
if x = O.
62
CHAPITRE 2. THE NETWORK CASE
2.2.3
Definition of viscosity solutions
We now introduce the definition of a viscosity solution of
λu(x) +
{−Du(x, ζ) − ξ} = 0
sup
(ζ,ξ)∈FL(x)
in G .
(2.2.1)
Definition 2.2.2.
• An upper semi-continuous function u : G → R is a subsolution of (2.2.1) in G if
for any x ∈ G , any φ ∈ R(G ) s.t. u − φ has a local maximum point at x, then
λu(x) +
sup
{−Dφ(x, ζ) − ξ} ≤ 0;
(2.2.2)
(ζ,ξ)∈FL(x)
• A lower semi-continuous function u : G → R is a supersolution of (2.2.1) if for any x ∈ G , any
φ ∈ R(G ) s.t. u − φ has a local minimum point at x, then
λu(x) +
sup
{−Dφ(x, ζ) − ξ} ≥ 0;
(2.2.3)
(ζ,ξ)∈FL(x)
• A continuous function u : G → R is a viscosity solution of (2.2.1) in G if it is both a viscosity
subsolution and a viscosity supersolution of (2.2.1) in G .
Remark 2.2.1. At x ∈ Ji \{O}, the notion of sub, respectively super-solution in Definition 2.2.2 is equivalent
to the standard definition of viscosity sub, respectively super-solution of
λu(x) + sup{− f i (x, a) · Du(x) − `i (x, a)} = 0.
a∈Ai
2.2.4
Hamiltonians
We define the Hamiltonians H i : Ji × R → R by
H i (x, p) = max(−p f i (x, a) − `i (x, a))
a∈Ai
(2.2.4)
and the Hamiltonian HO : RN → R by
HO (p1 , . . . , pN ) = max
i=1,...,N a∈Ai
max
(−pi f i (O, a) − `i (O, a)).
s.t. f i (O,a)≥0
(2.2.5)
We also define what may be called the tangential Hamiltonian at O by
HOT = − min
i=1,...,N a∈Ai
min
`i (O, a).
s.t. f i (O,a)=0
(2.2.6)
Thanks to the definitions of FLi (x) (in particular of FL+i (O)), the continuity properties of the data and
the compactness of Ai , one easily notes that the following definition is equivalent to Definition 2.2.2.
Definition 2.2.3.
• An upper semi-continuous function u : G → R is a subsolution of (2.2.1) in G if
for any x ∈ G , any φ ∈ R(G ) s.t. u − φ has a local maximum point at x, then
dφ
λu(x) + H i (x, d x i (x)) ≤ 0
λu(O) +
dφ
dφ
HO ( d x 1 (O), . . . , d x N
if x ∈ Ji \{O},
(O)) ≤ 0.
(2.2.7)
• A lower semi-continuous function u : G → R is a supersolution of (2.2.1) if for any x ∈ G , any
φ ∈ R(G ) s.t. u − φ has a local minimum point at x, then
dφ
λu(x) + H i (x, d x i (x)) ≥ 0
λu(O) +
dφ
dφ
HO ( d x 1 (O), . . . , d x N
if x ∈ Ji \{O},
(O)) ≥ 0.
(2.2.8)
2.2. THE HAMILTON-JACOBI EQUATION
63
The Hamiltonian H i are continuous with respect to x ∈ Ji , convex with respect to p. Moreover
p 7→ H i (O, p) is coercive, i.e. lim|p|→+∞ H i (O, p) = +∞ from the controlability assumption [H3].
Following Imbert-Monneau [IM14a], we introduce the nonempty compact interval P0i
P0i = {p0i ∈ R s.t. H i (O, p0i ) = min H i (O, p)}.
(2.2.9)
p∈R
Lemma 2.2.1. Assume [H0],[H1],[H2] and [H3], then
1. p0i ∈ P0i if and only if there exists a∗ ∈ Ai such that f i (O, a∗ ) = 0 and H i (O, p0i ) = −p0i f i (O, a∗ ) −
`i (O, a∗ ) = −`i (O, a∗ )
2.
min H i (O, p) = −
p∈R
a∈Ai
min
`i (O, a)
s.t. f i (O,a)=0
(2.2.10)
3. For all p ∈ R, if p ≥ p0i for some p0i ∈ P0i then
a∈Ai
max
(−p f i (O, a) − `i (O, a)) = min H i (O, q) = −
min
`i (O, a).
q∈R
a∈Ai s.t. f i (O,a)=0
s.t. fi (O,a)≥0
Proof. The Hamiltonian H i reaches its minimum at p0i if and only if 0 ∈ ∂ H i (O, p0i ). The subdifferential
of H i (O, ·) at p0i is characterized by
∂ H i (O, p0i ) = co{− f i (O, a); a ∈ Ai s.t. H i (O, p0i ) = −p0i f i (O, a) − `i (O, a)},
see [Val69]. But from [H2],
{( f i (O, a), `i (O, a)); a ∈ Ai s.t. H i (O, p0i ) = −p0i f i (O, a) − `i (O, a)}
is compact and convex. Hence,
∂ H i (O, p0i ) = {− f i (O, a); a ∈ Ai s.t. H i (O, p0i ) = −p0i f i (O, a) − `i (O, a)}.
Therefore, 0 ∈ ∂ H i (O, p0i ) if and only if there exists a∗ ∈ Ai such that f i (O, a∗ ) = 0 and H i (O, p0i ) =
−`i (O, a∗ ). We have proved point 1.
Point 2 is a direct consequence of point 1.
If p is greater than or equal to some p0i ∈ P0i , then
max
(−p f i (O, a) − `i (O, a)) ≤
a∈Ai : f i (O,a)≥0
max
(−p0i f i (O, a) − `i (O, a)) = H i (O, p0i )
a∈Ai : f i (O,a)≥0
where the last identity comes from point 1.
On the other hand,
max
a∈Ai : f i (O,a)≥0
(−p f i (O, a) − `i (O, a)) ≥ −
min
a∈Ai : f i (O,a)=0
`i (O, a).
Point 3 is obtained by combining the two previous observations and point 2.
Remark 2.2.2. It can also be proved that p ≤ max(q ∈ P0i ) if and only if
H i (O, p) =
max
(−p f i (O, a) − `i (O, a)).
a∈Ai : f i (O,a)≥0
In Figure 2.1, we give an example for the graphs of p 7→ H i (O, p) and of
p 7→ H i+ (O, p) ≡ maxa∈Ai : f i (O,a)≥0 (−p f i (O, a) − `i (O, a)), and the related interval P0i .
64
CHAPITRE 2. THE NETWORK CASE
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
−3
−2
−1
0
1
2
4
3
5
Figure 2.1: The graphs of the Hamiltonian p 7→ H i (O, p) (with the circles) and of p 7→ H i+ (O, p) ≡
maxa∈Ai : f i (O,a)≥0 (−p f i (O, a) − `i (O, a)) (with the signs +). In the example, P0i = [0, 2].
2.2.5
Existence
Theorem 2.2.1. Assume [H0],[H1],[H2] and [H3]. The value function v defined in (2.1.5) is a bounded
viscosity solution of (2.2.1) in G .
The proof of Theorem 2.2.1 is made in several steps, namely Proposition 2.2.1 and Lemmas 2.2.2
and 2.2.3 below: the first step consists of proving that the value function is a viscosity solution of a
Hamilton-Jacobi equation with a more general definition of the Hamiltonian: for that, we introduce
larger relaxed vector fields: for x ∈ G ,


t n → 0+ and
∃( y x,n , αn )n∈N ,


Z tn
1
f˜(x) = η ∈ Rd : ( y x,n , αn ) ∈ T x , s.t.
lim
f ( y x,n (t), αn (t))d t = η 

∃(t n )n∈N
n→∞ t
n
0
and
e
f`(x)
=

t n → 0+ ,

Z tn



1
∃( y x,n , αn )n∈N ,

lim
f ( y x,n (t), αn (t))d t = η,
(η, µ) ∈ Rd × R : ( y x,n , αn ) ∈ T x , s.t. n→∞ t n 0
Z

tn

∃(t n )n∈N

lim 1

`( y x,n (t), αn (t))d t = µ

n→∞
tn 0






.





Proposition 2.2.1. Assume [H0],[H1],[H2] and [H3]. The value function v defined in (2.1.5) is a viscosity
solution of
λu(x) + sup {−Du(x, ζ) − ξ} = 0 in G ,
(2.2.11)
e
(ζ,ξ)∈ f`(x)
e
where the definition of viscosity solution is exactly the same as Definition 2.2.2, replacing FL(x) with f`(x).
Proof. See [ACCT13].
For all φ ∈ R(G ), it is clear that if x ∈ Ji \{O}, then H i (x, Dφ) = sup(ζ,ξ)∈ f`(x)
{−Dφ(x, ζ) − ξ}.
e
We are left with comparing sup(ζ,ξ)∈FL(O) {−Dφ(O, ζ) − ξ} and sup(ζ,ξ)∈ f`(O)
{−Dφ(O, ζ) − ξ}. The two
e
quantities are the same. This is a consequence of the following lemma
Lemma 2.2.2.
¨
e
f`(O)
=
[
i=1,...,N
co
FL+i (O) ∪
[€
«
Š
FL j (O) ∩ ({0} × R) .
j6=i
Proof. The proof being a bit long, we postpone it to the appendix.
2.3. PROPERTIES OF VISCOSITY SUB AND SUPERSOLUTIONS
65
Lemma 2.2.3. Assume [H0],[H1],[H2] and [H3]. For any function φ in R(G ),
{−Dφ(O, ζ) − ξ} =
sup
e
(ζ,ξ)∈ f`(O)
max {−Dφ(O, ζ) − ξ}.
(ζ,ξ)∈FL(O)
(2.2.12)
e
Proof. It was proved in [ACCT13] that FL(O) ⊂ f`(O).
Hence
max {−Dφ(O, ζ) − ξ} ≤
(ζ,ξ)∈FL(O)
{−Dφ(O, ζ) − ξ}.
sup
e
(ζ,ξ)∈ f`(O)
From the piecewise linearity of the function (ζ, µ) 7→ −Dφ(O, ζ) − µ, we infer that
(ζ,µ)∈co
= max
FL+
i (O)∪
sup€
S
j6=i
max
Š(−Dφ(O, ζ) − µ)
FL j (O)∩({0}×R)
(−Dφ(O, ζ) − µ), max
(ζ,µ)∈FL+
i (O)
max
j6=i (0,µ)∈FL j (O)
−µ
≤ max j=1,...,N max(ζ,µ)∈FL+j (O) −Dφ(O, ζ) − µ) = max(ζ,ξ)∈FL(O) {−Dφ(O, ζ) − ξ}.
We conclude by using Lemma 2.2.2.
2.3
Properties of viscosity sub and supersolutions
In this part, we study sub and supersolutions of (2.2.1), transposing ideas coming from Barles-BrianiChasseigne [BBC13, BBC14] to the present context.
Lemma 2.3.1. Assume [H0],[H1],[H2] and [H3]. Let R be a positive real number such that for all i =
1, . . . , N and x ∈ B(O, R) ∩ Ji
δ δ
[− ei , ei ] ⊂ Fi (x).
2
2
For any bounded viscosity subsolution u of (2.2.1), there exists a constant C ∗ > 0 such that u is a viscosity
subsolution of
|Du(x)| ≤ C ∗ in B(O, R) ∩ G ,
i.e. for any x ∈ B(O, R) ∩ G and φ ∈ R(G ) such that u − φ has a local maximum point at x,
dφ
(x)| ≤ C ∗
d xi
dφ
min
(O) ≥ −C ∗
i d xi
|
if x ∈ B(O, R) ∩ Ji \{O}
(2.3.1)
if x = O.
(2.3.2)
Proof. Let Mu (resp M` ) be an upper bound on |u| (resp. ` j for all j = 1, . . . , N ). The viscosity inequality
(2.2.7) yields that
dφ
(x))
d xi
dφ
dφ
HO (
(O), . . . ,
(O))
d x1
d xN
H i (x,
≤ λMu
≤ λMu
if x ∈ B(O, R) ∩ Ji \{O},
(2.3.3)
if x = O.
(2.3.4)
From the controlability in B(O, R) ∩ Ji , we see that H i is coercive with respect to its second argument
uniformly in x ∈ B(O, R) ∩ Ji , and more precisely that H i (x, p) ≥ δ2 |p| − M` .
λM +M
dφ
Thus, from (2.3.3), there exists a constant C ∗ = 2 uδ ` such that | d x i (x)| ≤ C ∗ if x ∈ B(O, R)∩ Ji \{O}.
If x = O, we use the fact that H i+ (O, p) ≥
that mini
dφ
d x i (O)
δ
2
max(0, −p)− M` . The viscosity inequality (2.3.4) then yields
≥ −C ∗ .
Lemma 2.3.2. Assume [H0],[H1],[H2] and [H3]. There exist r > 0 such that any bounded viscosity
subsolution u of (2.2.1) is Lipschitz continuous in B(O, r) ∩ G .
66
CHAPITRE 2. THE NETWORK CASE
Proof. We adapt the proof of H.Ishii, see [Ish13].
Take R as in Lemma 2.3.1, fix z ∈ B(O, R) ∩ G and set r = (R − |z|)/4. Fix any y ∈ G such that
d( y, z) < r. It can be checked that for any x ∈ G , if d(x, y) < 3r then d(x, O) < R. Choose a function
f ∈ C 1 ([0, 3r)) such that f (t) = t in [0, 2r] and f 0 (t) ≥ 1 for all t ∈ [0, 3r) and lim t→3r f (t) = +∞.
Fix any " > 0. We are going to show that
u(x) ≤ u( y) + (C ∗ + ") f (d(x, y)),
∀x ∈ G such that d(x, y) < 3r,
(2.3.5)
where C ∗ is the constant in Lemma 2.3.1.
Let us proceed by contradiction. Assume that (2.3.5) is not true. According to the properties of
f , the function x 7→ u(x) − u( y) − (C ∗ + ") f (d(x, y)) admits a maximum ξ ∈ B( y, 3r) ∩ G . However,
from the fact (2.3.5) is not true, we deduce that ξ 6= y. Hence, it is possible to modify the function
ψ : G → R, x 7→ (C ? + ") f (d(x, y)) away from a neighborhood of ξ and obtain an admissible test
function that we use in the viscosity inequality satisfied by u; from (2.3.1) and (2.3.2) in Lemma 2.3.1
and from explicit calculatiobns concerning the derivatives of d(x, y) at the point ξ, we obtain that
(C ∗ + ") f 0 (d(ξ, y)) ≤ C ∗ ,
which leads to a contradiction.
If d(x, z) < r then d(x, y) < 2r and f (d(x, y)) = d(x, y). In this case, (2.3.5) yields that
u(x) ≤ u( y) + (C ∗ + ")d(x, y),
∀x, y ∈ G s.t d(x, z) < r, d( y, z) < r.
By symmetry, we get
|u(x) − u( y)| ≤ (C ∗ + ")d(x, y),
∀x, y ∈ G s.t d(x, z) < r, d( y, z) < r,
and by letting " tend to zero:
|u(x) − u( y)| ≤ C ∗ d(x, y),
∀x, y ∈ G s.t d(x, z) < r, d( y, z) < r.
(2.3.6)
Now, for two arbitrary points x, y in G ∩ B(O, R), we take r = 14 min(R − |x|, R − | y|) and choose a
finite sequence (z j ) j=1,...,M ∈ G belonging to the segment [x, y] if x and y belong to some Ji or to
[O, x] ∪ [O, y] in the opposite case, and such that z1 = x, z M = y and d(zi , zi+1 ) < r for all i =
P M −1
1, . . . , M − 1 and i=1 d(zi , zi+1 ) = d(x, y). From (2.3.6), we get that
|u(x) − u( y)| ≤ C ∗ d(x, y),
∀x, y ∈ G ∩ B(O, R).
Lemma 2.3.3. Assume [H0],[H1],[H2] and [H3]. Any bounded viscosity subsolution u of (2.2.1) is such
that
λu(O) ≤ −HOT .
(2.3.7)
Proof. Since, from Lemma 2.3.2, u is Lipschitz continuous in a neighborhood of O, we know that there
exists a test-function φ in R(G ) which touches u from above at O. Since u is a subsolution of (2.2.1),
dφ
dφ
we see that λu(O) + HO ( d x 1 (O), . . . , d x N (O)) ≤ 0, which implies that λu(O) + HOT ≤ 0.
Remark 2.3.1. It is interesting to note that in [BBC13] and [BBC14], a condition similar to (2.3.7) is
introduced to characterize a particular viscosity solution of the transmission problem studied there among
all the possible solutions in the sense of Ishii, (this condition is not satisfied by all subsolutions).
In the present context, the fact that (2.3.7) is automatically satisfied by subsolutions seems to be linked to
the richness of the space R(G ): for any Lipschitz function u defined in a neighborhood of O, there exists
φ ∈ R(G ) such that u − φ has a maximum at O.
The following lemma can be found in [BBC13, BBC14] in a different context:
2.3. PROPERTIES OF VISCOSITY SUB AND SUPERSOLUTIONS
67
Lemma 2.3.4. Let v : G → R be a viscosity supersolution of (2.2.1) in G . Then if x ∈ Ji \{0}, we have for
all t > 0,
‚Z t∧θ
Œ
i
`i ( y xi (s), αi (s))e−λs ds + v( y xi (t ∧ θi ))e−λ(t∧θi ) ,
v(x) ≥ inf
αi (·),θi
(2.3.8)
0
where αi ∈ L ∞ (0, ∞; Ai ), y xi is the solution of y xi (t) = x +
€R t
0
Š
f i ( y xi (s), αi (s))ds ei and θi is such that
y xi (θi ) = 0 and θi lies in [τi , τ̄i ], where τi is the exit time of y xi from Ji \{O} and τ̄i is the exit time of y xi
from Ji .
Proof. See [BBC13] for the detailed proof. We restrict ourselves to mentioning that the proof of (2.3.8)
uses the results of Blanc [Bla97,Bla01] on the minimal supersolution of exit time control problems.
Remark 2.3.2. Note that comparison results of Barles-Perthame [BP90] imply the following suboptimality
principle for subsolutions that will not be needed in the sequel: let w be a continuous viscosity subsolution
of (2.2.1) in G . If x ∈ Ji \{0}, we have for all t > 0,
‚Z
t∧θi
`i ( y xi (s), αi (s))e−λs ds
w(x) ≤ inf sup
αi (·) θ
i
+
w( y xi (t
∧ θi ))e
−λ(t∧θi )
Œ
,
(2.3.9)
0
where αi ∈ L ∞ (0, ∞; Ai ), y xi is the solution of y xi (t) = x +
€R t
0
Š
f i ( y xi (s), αi (s))ds ei and θi is such that
y xi (θi ) = 0 and θi lies in [τi , τ̄i ], where τi is the exit time of y xi from Ji \{O} and τ̄i is the exit time of y xi
from Ji .
The following theorem is reminiscent of Theorem 3.3 in [BBC13]:
Theorem 2.3.1. Assume [H0],[H1],[H2] and [H3]. Let r > 0 be given by Lemma 2.3.2. Let v : G → R
be a viscosity supersolution of (2.2.1), bounded from below by −c|x| − C for two positive numbers c and
C. Either [A] or [B] below is true:
[A] There exists a sequence (ηk )k∈N of positive real numbers such that limk→+∞ ηk = η > 0, an index
i ∈ {1, . . . , N } and a sequence x k ∈ Ji such that x k ∈ Ji \ {O} and limk→+∞ x k = O satisfying the
following: for any k ∈ N, there exists a control law αki such that the corresponding trajectory y x k
remains in Ji ∩ B(O, r) in the time interval [0, ηk ], i.e. y x k (s) ∈ Ji ∩ B(O, r) for all s ∈ [0, ηk ], and
is such that
Z ηk
`i ( y x k (s), αki (s))e−λs ds + v( y x k (η))e−ληk
v(x k ) ≥
(2.3.10)
0
[B]
λv(O) + HOT ≥ 0.
(2.3.11)
Proof. Let us assume that [B] does not hold.
For any i in {1, . . . , N }, take for example
qi = min p0i ,
p0i ∈P0i
and q = (q1 , . . . , qN ). From Lemma 2.2.1,
HO (q) = HOT .
Consider the function
(2.3.12)
|x|2
if x ∈ Ji .
"2
Standard arguments show that this function reaches its minimum near O and any sequence of such
minimum points x " converges to O and that v(x " ) converges to v(O).
It is not possible that x " be O, because since v is a viscosity supersolution of (2.2.1), we would have
that
λv(O) + HO (q) ≥ 0,
v(x) − qi |x| +
68
CHAPITRE 2. THE NETWORK CASE
and therefore λv(O) + HOT ≥ 0, which is a contradiction since [B] does not hold.
Therefore, there exists i ∈ {1, . . . , N } such that, up to the extraction of a subsequence, x " ∈ (Ji \{O}) ∩
B(O, 2r ), for all ". We can therefore apply Lemma 2.3.4: for any t > 0,
t∧θi
‚Z
v(x " ) ≥ inf
αi (·),θi
`i ( y xi (s), αi (s))e−λs ds
"
0
+
v( y xi (t
"
∧ θi ))e
−λ(t∧θi )
Œ
,
(2.3.13)
Š
€R t
where y xi is the solution of y xi (t) = x + 0 f i ( y xi (s), αi (s))ds ei .
Take t = 4Mr f for example. From [H0] and [H2], the minimum in (2.3.13) is reached for some αi," and
θi," > 0, see [BBC13] :
Z
v(x " ) ≥
t∧θi,"
`i ( y xi (s), αi," (s))e−λs ds + v( y xi (t ∧ θi," ))e−λ(1∧θi," ) .
"
0
"
(2.3.14)
If there exists a subsequence (still called θi," ) such that lim"→0 θi," = θ > 0, then we obtain [A] with
η" = t ∧ θi," and η = t ∧ θ . Note that since t = 4Mr f and x " ∈ (Ji \{O}) ∩ B(O, 2r ), we deduce from
Assumption [H0] that y x " (s) ∈ Ji ∩ B(O, r) for all s ∈ [0, η" ].
Assume by contradiction that [A] does not hold: then, from the latter argument, lim"→0 θi," = 0.
Since x " is a minimum of v(x) − qi |x| +
Z
|x|2
"2 ,
we deduce from (2.3.14) that
θi,"
|x " |2
,
"2
(2.3.15)
`i ( y xi (s), αi," (s))e−λs ds + v( y xi (θi," ))(e−λθi," − 1) − qi |x " |.
(2.3.16)
`i ( y xi (s), αi," (s))e−λs ds + v( y xi (θi," ))(e−λθi," − 1) − qi |x " | +
0≥
"
"
0
and therefore
θi,"
Z
0≥
"
0
"
We can write (2.3.16) as
Z
θi,"
0≤
0
€
Š
−`i ( y xi (s), αi," (s))e−λs − qi f i ( y xi (s), αi," (s)) ds − v( y xi (θi," ))(e−λθi," − 1).
"
"
"
(2.3.17)
Dividing by θi," and letting " tend to 0, we obtain that λv(O) + H i (O, qi ) ≥ 0. This implies that λv(O) +
HOT ≥ 0, which is a contradiction since [B] does not hold.
Lemma 2.3.5. Assume [H0],[H1],[H2] and [H3]. Let r > 0 be given by Lemma 2.3.2: any bounded
subsolution of (2.2.1) is Lipschitz continuous in B(O, r) ∩ G . Consider
€R ti ∈ {1, · · · , N }, x Š∈ (Ji \ {O}) ∩
∞
B(O, r), αi ∈ L (0, ∞; Ai ). Let η > 0 be such that y x (t) = x + 0 f i ( y x (s), αi (s))ds ei belongs to
Ji ∩ B(O, r) for any t ∈ [0, η]. For any bounded viscosity subsolution v of (2.2.1),
Zη
`i ( y x (t), αi (t))e−λt d t + v( y x (η))e−λη .
v(x) ≤
(2.3.18)
0
Proof. Since v is Lipschitz continuous in B(O, r) ∩ Ji , the function t 7→ v( y x (t))e−λt is Lipschitz continuous in [0, η]. Let us define the sets KO = {t ∈ [0, η] : y x (t) = O} and KOc = [0, η]\KO . It is clear that
KO is closed and that KOc is an open subset of [0, η]. We first observe that, from Stampacchia’s theorem,
Z
η
0
1KO (t)
d
v( y x (t))e−λt d t = −λv(O)
dt
Z
η
1KO (t)e−λt d t.
0
Therefore, we deduce from Lemma 2.3.3 that
Zη
Zη
Zη
Zη
d
1KO (t)
v( y x (t))e−λt ) d t ≥ HOT
1KO (t)d t ≥ −
`i (O, αi (t))1KO (t)d t = −
`i ( y x (t), αi (t))1KO (t)d t.
dt
0
0
0
0
(2.3.19)
2.4. COMPARISON PRINCIPLE AND UNIQUENESS
69
On the other hand, since KOc is an open subset
S of [0, η], there exists a countable family of disjoint
intervals (ω j ) j∈J , ω j ⊂ [0, η] such that KOc = j∈J ω j . Let a j < b j be the lower and upper endpoints
of ω̄ j . We can assume that [a j , b j ] ∩ [ak , bk ] = ; if j 6= k. From a classical suboptimality principle,
see [BCD97, Theorem III.2.33], we see that for any j ∈ J,
Z bj
v( y x (b j ))e−λb j − v( y x (a j ))e−λa j ≥ −
`i ( y x (t), αi (t))e−λt d t.
aj
Noting that
v( y x (b j ))e
−λb j
− v( y x (a j ))e
−λa j
=
Z
η
0
d
v( y x (t))e−λt 1(a j ,b j ) (t)d t,
dt
and summing over j ∈ J, we obtain that
Zη
Zη
d
−λt
v( y x (t))e ) d t ≥ −
`i ( y x (t), αi (t))1KOc (t)d t.
1KOc (t)
dt
0
0
(2.3.20)
We get (2.3.18) by summing (2.3.19) and (2.3.20).
2.4
Comparison principle and Uniqueness
Theorem 2.4.1. Assume [H0],[H1],[H2] and [H3]. Let u : G → R be a bounded viscosity subsolution of
(2.2.1), and v : G → R be a bounded viscosity supersolution of (2.2.1). Then u ≤ v in G .
Proof. It is a simple matter to check that there exists a positive real number M such that the function
ψ(x) = −|x|2 − M is a viscosity subsolution of (2.2.1). For 0 < µ < 1, µ close to 1, the function
uµ = µu + (1 − µ)ψ is a viscosity subsolution of (2.2.1), which tends to −∞ as |x| tends to +∞. Let
Mµ be the maximal value of uµ − v which is reached at some point x̄ µ .
We want to prove that Mµ ≤ 0.
1. If x̄ µ 6= O, then we introduce the function uµ (x) − v(x) − d 2 (x, x̄ µ ), which has a strict maximum
at x̄ µ , and we double the variables, i.e. for 0 < " 1, we consider
uµ (x) − v( y) − d 2 (x, x̄ µ ) −
d 2 (x, y)
.
"2
Classical arguments then lead to the conclusion that uµ (x̄ µ ) − v(x̄ µ ) ≤ 0, thus Mµ ≤ 0.
2. If x̄ µ = O. We use Theorem 2.3.1; we have two possible cases:
[B] λv(O) ≥ −HOT .
From Lemma 2.3.3, λu(O) + HOT ≤ 0. Therefore, we obtain that uµ (O) ≤ v(O), thus Mµ ≤ 0.
[A] With the notations of Theorem 2.3.1, we have that
Z ηk
`i ( y x k (s), αki (s))e−λs ds + v( y x k (ηk ))e−ληk .
v(x k ) ≥
0
Moreover, since y x k (s) ∈ Ji ∩ B(O, r) for all s ∈ [0, ηk ], Lemma 2.3.5 can be applied and
yields that
Z ηk
`i ( y x k (s), αki (s))e−λs ds + uµ ( y x k (ηk ))e−ληk .
uµ (x k ) ≤
0
Therefore
uµ (x k ) − v(x k ) ≤ (uµ ( y x k (ηk )) − v( y x k (ηk )))e−ληk .
Letting k tend to +∞, we find that Mµ ≤ Mµ e−λη , which implies that Mµ ≤ 0
We conclude by letting µ tend to 1.
Corollary 2.4.1. Assume [H0],[H1],[H2] and [H3]. The value function u of the optimal control problem
(2.1.5) is the unique bounded viscosity solution of (2.2.1).
70
CHAPITRE 2. THE NETWORK CASE
2.5
Stability
We now study the stability of sub and super solutions with respect to the uniform convergence of the
costs and dynamics.
2.5.1
Assumptions
We consider a family (indexed by " ∈ [0, 1]) of optimal control problems on the network whose dynamics and costs are denoted ( f i" , `"i ) for i = 1, . . . , N . As above, A is a metric space (one can take A = Rm )
and for i = 1, . . . , N , Ai are nonempty disjoint compact subsets of A. Hereafter, we suppose that the
following properties hold uniformly with respect to ":
[H0" ] The functions f i" : Ji × Ai → R are continuous and bounded uniformly w.r.t. " ∈ [0, 1]; in
particular, there exists M > 0 such that | f i" (x, a)| ≤ M for any " ∈ [0, 1], i = 1, . . . , N , x ∈ Ji ,
a ∈ Ai . Moreover, there exists L > 0 such that for any ", i, x, y in Ji and a ∈ Ai ,
| f i" (x, a) − f i" ( y, a)| ≤ L|x − y|.
We will use the notation Fi" (x) for the set { f i" (x, a)ei , a ∈ Ai }.
[H1" ] For i = 1, . . . , N , the functions `"i : Ji × Ai → R are continuous and bounded uniformly w.r.t.
" ∈ [0, 1]; we may assume that |`"i (x, a)| ≤ M for any " ∈ [0, 1], i = 1, . . . , N , x ∈ Ji , a ∈ Ai with
the same constant M as above. There is a modulus of continuity ωi such that for all " ∈ [0, 1],
x, y in Ji and a ∈ Ai , |`"i (x, a) − `"i ( y, a)| ≤ ωi (|x − y|).
[H2" ] For i = 1, . . . , N , x ∈ Ji , the non empty and closed set
FL"i (x) ≡ {( f i" (x, a)ei , `"i (x, a)), a ∈ Ai }
is convex.
[H3" ] There is a real number δ > 0 such that for any " ∈ [0, 1], i = 1, . . . , N ,
[−δei , δei ] ⊂ Fi" (O).
We also assume the local uniform convergence of f i" to f i0 and `"i to `0i as " → 0: for all i = 1, . . . , N
and R > 0,
[H4" ]
lim
max
| f i" (x, a) − f i0 (x, a)| = 0.
lim
max
|`"i (x, a) − `0i (x, a)| = 0.
"→0 x∈B(O,R),a∈Ai
[H5" ]
"→0 x∈B(O,R),a∈Ai
2.5.2
Convergence of the Hamiltonian at the vertex as " → 0
Lemma 2.5.1. For " fixed in [0, 1] and i ∈ {1, . . . , N }, let a∗ ∈ Ai be such that f i" (O, a∗ ) ≥ 0. There exists
a sequence an∗ ∈ Ai such that
f i" (O, an∗ )
≥
| f i" (O, an∗ ) − f i" (O, a∗ )|
≤
|`"i (O, an∗ ) − `"i (O, a∗ )|
≤
δ
> 0,
n
2M
,
n
2M
.
n
(2.5.1)
(2.5.2)
(2.5.3)
2.5. STABILITY
71
Proof. From [H3" ] there exists aδ ∈ Ai such that f i" (O, aδ ) = δ. From [H2" ],
λ( f i" (O, aδ ), `"i (O, aδ )) + (1 − λ)( f i" (O, a∗ ), `"i (O, a∗ )) ∈ FL"i (O)
for any λ ∈ [0, 1]. In particular, for λ = 1n , there exists an∗ ∈ Ai such that
1 "
1
( f (O, aδ ), `"i (O, aδ )) + (1 − )( f i" (O, a∗ ), `"i (O, a∗ )) = ( f i" (O, an∗ ), `"i (O, an∗ ))
n i
n
which yields (2.5.1). The statements (2.5.2) (2.5.3) follow from [H0" ] and [H1" ].
Corollary 2.5.1. For any " ∈ [0, 1], i ∈ {1, . . . , N } and pi ∈ R,
a∈Ai
max
(−pi f i" (O, a) − `"i (O, a)) =
sup
(−pi f i" (O, a) − `"i (O, a)).
"
s.t. f i" (O,a)≥0
a∈Ai s.t. f i (O,a)>0
(2.5.4)
As in the previous sections, we define the Hamiltonians
H i" (x, p)
=
HO" (p1 , . . . , pN )
=
With
"
HO,i
(pi ) =
max(−p f i" (x, a) − `"i (x, a)),
(2.5.5)
a∈Ai
max
i=1,...,N a∈Ai
a∈Ai
max
(−pi f i" (O, a) − `"i (O, a)).
s.t. f i" (O,a)≥0
max
(−pi f i" (O, a) − `"i (O, a)),
s.t. fi" (O,a)≥0
(2.5.6)
(2.5.7)
"
we can write HO" (p1 , . . . , pN ) = maxi=1,...,N HO,i
(pi ). Finally, we define
T,"
HO = − min
i=1,...,N a∈Ai
Proposition 2.5.1. For any p ∈ RN ,
min
`" (O, a).
s.t. f i" (x,a)=0 i
lim HO" (p) = HO0 (p).
(2.5.9)
lim sup HO" (p) ≤ HO0 (p).
(2.5.10)
"→0+
Proof. Let us first prove that
(2.5.8)
"→0+
For any i ∈ {1, . . . , N }, let (a" )" be a family of points in Ai such that f i" (O, a" ) ≥ 0. Up to the extraction
of subsequence, we can assume that there exists a0 ∈ Ai such that lim+ a" = a0 . Then f i0 (O, a0 ) ≥ 0
"→0
and
(−pi f i" (O, a" ) − `"i (O, a" )) = (−pi f i0 (O, a0 ) − `0i (O, a0 )) + o(1).
This implies that
lim sup
"→0
a∈Ai
max
(−pi f i" (O, a" ) − `"i (O, a" )) ≤
max
(−pi f i0 (O, a0 ) − `0i (O, a0 ))
s.t f i" (O,a)≥0
a∈Ai s.t f i0 (O,a)≥0
i.e. (2.5.10).
We are left with proving that
lim inf
HO" (p) ≥ HO0 (p).
+
(2.5.11)
"→0
For a positive integer n, call A"i,n,δ the set
A"i,n,δ = {a ∈ Ai s.t. f i" (O, a) ≥
δ
}.
n
The set A0i,n,δ is compact and from [H4" ], there exists "¯n such that for any " ≤ "¯n ,
A0i,n,δ ⊂ A"i,2n,δ ⊂ {a ∈ Ai s.t. f i" (O, a) ≥ 0}.
This implies that
max (−pi f i0 (O, a) − `0i (O, a)) ≤
a∈A0i,n,δ
a∈Ai
max
(−pi f i" (O, a) − `"i (O, a)) + o(1)
s.t. f i" (O,a)≥0
72
CHAPITRE 2. THE NETWORK CASE
and letting " → 0
max (−pi f i0 (O, a) − `0i (O, a))
max
i=1,...,N a∈A0
i,n,δ
≤ lim inf
max
+
"→0
i=1,...,N a∈Ai
(−pi f i" (O, a) − `"i (O, a)).
max
s.t. f i" (O,a)≥0
Therefore, for any positive integer n,
max
max (−pi f i0 (O, a) − `0i (O, a)) ≤ lim inf
HO" (p)
+
"→0
i=1,...,N a∈A0
i,n,δ
(2.5.12)
Consider now a0 ∈ Ai such that
−pi f i0 (O, a0 ) − `0i (O, a0 ) = HO0 (p)
= max
j=1,...,N a∈A j
max
(−p j f j0 (O, a) − `0j (O, a)).
s.t. f i0 (O,a)≥0
From Lemma 2.5.1, there exists a sequence (an0 )n>0 such that an0 ∈ Ai,n,δ and
lim (−pi f i0 (O, an0 ) − `0i (O, an0 )) = (−pi f i0 (O, a0 ) − `0i (O, a0 )) = HO0 (p).
n→∞
From (2.5.12),
(−pi f i0 (O, an0 ) − `0i (O, an0 )) ≤ lim inf
HO" (p)
+
"→0
(2.5.13)
which yields (2.5.11) by letting n → ∞.
Remark 2.5.1. Note that for proving Proposition 2.5.1, only [H20 ] , [H30 ] are needed, (in addition to
[H0" ], [H1" ], [H4" ] and [H5" ]).
Remark 2.5.2. It is possible to prove under the hypotheses of the Proposition 2.5.1 that for any pi ∈ R,
0
"
(pi ).
(pi ) = HO,i
lim HO,i
(2.5.14)
"→0
The proof is very much like that of Proposition 2.5.1.
2.5.3
Convergence of the sub or super solutions as " → 0
We consider the family of Hamilton-Jacobi equations depending on the parameter ":
λu(x) +
λu(x) +
sup
{−Du(x, ζ) − ξ}
=
0
in G ,
(2.5.15)
sup
{−Du(x, ζ) − ξ}
=
0
in G .
(2.5.16)
(ζ,ξ)∈FL" (x)
(ζ,ξ)∈FL0 (x)
Theorem 2.5.1. Let u" be a sequence of uniformly Lipschitz subsolutions of (2.5.15) converging to u0 as
" → 0 locally uniformly on G . Then u0 is a subsolution of (2.5.16).
Proof. Consider x 0 ∈ G and φ ∈ R(G ) such that x 0 is a strict local maximum point of u0 − φ; we wish
to prove that
dφ
(x 0 ))
d xi
dφ
dφ
λu0 (O) + HO0 (
(O), . . . ,
(O))
d x1
d xN
λu0 (x 0 ) + H i0 (x 0 ,
≤
0
if x 0 ∈ Ji \{O},
≤
0
if x 0 = O.
The proof is standard if x 0 6= O. Let us assume that x 0 = O. We have to prove that
λu0 (O) + max
i=1,...,N a∈Ai
dφ
max
(−
(O) f i0 (O, a) − `0i (O, a)) ≤ 0.
s.t. f i0 (O,a)≥0 d x i
(2.5.17)
2.5. STABILITY
73
Having fixed i ∈ {1 . . . N }, define
di ( y) =
§
0 if y ∈ Ji ,
| y| otherwise.
Let L̄ be an uniform bound of the Lipschitz constant of u" − φ. Take C = L̄ + 1.
The function y 7→ u0 ( y) − φ( y) − C di ( y) reaches a strict local maximum point at O, say in B(O, R).
Thanks to the local uniform convergence of u" , there exists a sequence of local maximum points y " in
B(O, R) of y 7→ u" ( y) − φ( y) − C di ( y) which converges to O as " → 0.
Moreover y " ∈ Ji , because if it was not the case, then
u" ( y " ) − φ( y " ) − u" (O) − φ(O) ≤ L̄| y " | = L̄di ( y " ),
would imply
u" ( y " ) − φ( y " ) − C di ( y " ) ≤ u" (O) − φ(O) − di ( y " ) < u" (O) − φ(O),
which would contradict the definition of y " .
Then, take y 7→ φ( y) + C di ( y) as a test function in the viscosity inequality satisfied by u" . We make
out two cases:
Case 1: y " ∈ Ji \{O}. We obtain
λu" ( y " ) + H i" ( y " ,
dφ "
( y )) ≤ 0,
d xi
and letting " → 0
λu0 (O) + H i0 (O,
dφ
(O)) ≤ 0.
d xi
(2.5.18)
Case 2: y " = O.
λu" (O) + max
j=1,...,N a∈A j
where p j =
dφ
d x j (O) +
λu" (O) +
a∈Ai
max
(−p j f j" (O, a) − `"j (O, a)) ≤ 0,
s.t. f j" (O,a)≥0
C if j 6= i and pi =
dφ
d x i (O).
Hence,
dφ
dφ
"
max
(−
(O) f i" (O, a) − `"i (O, a)) = λu" (O) + HO,i
(
(O)) ≤ 0.
d xi
s.t. f i" (O,a)≥0 d x i
From (2.5.14), we deduce that
0
λu0 (O) + HO,i
(
dφ
(O)) ≤ 0.
d xi
(2.5.19)
Summarizing, we have (2.5.19) in all cases, because (2.5.18) implies (2.5.19). We have proved (2.5.17).
Theorem 2.5.2. Let (u" )" be a sequence of supersolutions of (2.5.15) such that
• there exist a real number C > 0 s.t. for all " and x ∈ G , |u" (x)| ≤ C(1 + |x|)
• the sequence u" converges to u0 locally uniformly on G as " → 0.
Then u0 is a supersolution of (2.5.16).
Proof. Consider x 0 ∈ G and φ ∈ R(G ) such that x 0 is a strict local minimum point of u0 − φ; if x 0 6= O,
dφ
the proof that λu0 (x 0 ) + H i0 (x 0 , d x i (x 0 )) ≥ 0 is standard. We therefore focus on the case when x 0 = O.
We consider two cases:
74
CHAPITRE 2. THE NETWORK CASE
dφ
dφ
dφ
dφ
First case: for any i = 1, . . . , N , d x i (O) ≤ max(q : q ∈ P0i ) and H i0 (O, d x i (O)) = HO0 ( d x 1 (O), . . . , d x N (O)).
In this case, we can use the standard stability argument: there exists a sequence (x " ) such that
x " is a local minimum point of u" − φ and such that x " converges to O and u" (x " ) converges to
u0 (O). If for a subsequence "n , x "n = O, then the viscosity inequality is
"
λu"n (O) + HOn (
dφ
dφ
(O), . . . ,
(O)) ≥ 0
d x1
d xN
and by passing to the limit as n → ∞ thanks to Proposition 2.5.1,
λu0 (O) + HO0 (
dφ
dφ
(O), . . . ,
(O)) ≥ 0,
d x1
d xN
(2.5.20)
which is the desired viscosity inequality for u0 . If there does not exists such a subsequence, we
can assume that for a subsequence "n , x "n ∈ Ji \{O}. The viscosity inequality is
"
λu"n (x "n ) + H i n (x "n ,
dφ "n
(x ) ≥ 0,
d xi
and by passing to the limit as n → ∞,
λu0 (O) + H i0 (O,
dφ
(O)) ≥ 0.
d xi
dφ
dφ
dφ
Then (2.5.20) is obtained since H i0 (O, d x i (O)) = HO0 ( d x 1 (O), . . . , d x N (O)).
Second case: I 6= {1, . . . , N }, where I is the (possibly empty) set of indices i such that
P0i )
dφ
H i0 (O, d x i (O))
max(q : q ∈
and
function ψ ∈ R(G ) such that
=
dφ
dφ
HO0 ( d x 1 (O), . . . , d x N
dφ
d x i (O)
≤
(O)). It is always possible to find a
1. ψ(O) = φ(O)
dψ
dψ
dφ
dφ
2. HO0 ( d x 1 (O), . . . , d x N (O)) = HO0 ( d x 1 (O), . . . , d x N (O))
3. if i ∈ I , then ψ|Ji coincides with φ|Ji
dψ
dφ
d x i (O) < d x i (O)
dφ
dφ
HO0 ( d x 1 (O), . . . , d x N (O)).
4. if i 6∈ I , then
is such that
dψ
d x i (O)
dψ
≤ max(q : q ∈ P0i ) and H i0 (O, d x i (O)) =
Then, since ψ touches φ at O from below, O is still a strict minimum point of u0 − ψ, and for all
dψ
i, d x i (O) ≤ max(q : q ∈ P0i ) and
H i0 (O,
dψ
dψ
dψ
dφ
dφ
(O)) = HO0 (
(O), . . . ,
(O)) = HO0 (
(O), . . . ,
(O)).
d xi
d x1
d xN
d x1
d xN
(2.5.21)
We can apply the result proved in the first case to the function ψ, i.e.
λu0 (O) + HO0 (
dψ
dψ
(O), . . . ,
(O)) ≥ 0,
d x1
d xN
and we get (2.5.20) from (2.5.21).
2.6
Extension to a more general framework with an additional
cost at the junction
It is possible to extend all the results presented above to the case when there is an additional cost at the
junction. Such problems are also studied in [IM14a]. We keep the setting used above except that we
2.7. THE CASE OF A NETWORK
75
take into account an additional subset A0 of A (it is enough to suppose that A0 is a singleton and that it
is disjoint from the other sets Ai ), on which the running cost is the constant `0 . We define
M = (x, a); x ∈ G , a ∈ Ai if x ∈ Ji \{O}, and a ∈ ∪Ni=0 Ai if x = O ,
the dynamics
∀(x, a) ∈ M ,

 f i (x, a)ei
f i (O, a)ei
f (x, a) =
 0
if x ∈ Ji \{O},
if x = O and a ∈ Ai , i > 0,
if x = O and a ∈ A0 ,
and the running cost
∀(x, a) ∈ M ,

 `i (x, a)
`i (O, a)
`(x, a) =
 `
0
if x ∈ Ji \{O},
if x = O and a ∈ Ai , i > 0,
if x = O and a ∈ A0 .
The infinite horizon optimal control problem is then given by (2.1.5) and (2.1.4). We obtain that the
value function v is continuous in the same manner as above and that v is a viscosity solution of (2.2.1)
with the new definition of FL(x):
§
FLi (x)
if x belongs to the edge Ji \{O}
S
FL(x) =
{0, −`0 } ∪ i=1,...,N FL+i (O)
if x = O.
The viscosity sub and supersolutions can be also defined as in (2.2.7) and (2.2.8) with the new definition
of HO : RN → R:

‹
HO (p1 , . . . , pN ) = max −`0 , max
max
(−pi f i (O, a) − `i (O, a)) ,
i=1,...,N a∈Ai s.t. f i (O,a)≥0
and the definition of the constant HOT is modified accordingly:

‹
T
HO = − min `0 , min
min
`i (O, a) .
i=1,...,N a∈Ai s.t. f i (O,a)=0
With these new definitions, all the results proved in § 2.3, 2.4 and 2.5 hold with obvious modifications
of the proofs. In particular,
• a subsolution of the presently defined problem is also a subsolution of the former problem (without the additional cost) so it is Lipschitz continuous in a neighborhood of O, and Lemmas 2.3.2,
2.3.3 and 2.3.5 hold.
• The proofs of Lemma 2.3.4 and Theorem 2.3.1 are unchanged. In particular, with the choice of
q = (qi )i=1,...,N made in the proof of Theorem 2.3.1, we still have the identity HO (q) = HOT .
• The proof of the comparison principle is unchanged.
2.7
2.7.1
The case of a network
The geometrical setting and the optimal control problem
We consider a network in Rd with a finite number of edges and vertices. A network in Rd is a pair
(V , E ) where
i) V is a finite subset of Rd whose elements are said vertices
ii) E is a finite set of edges, which are either closed straight line segments between two vertices, or
a closed straight half-lines whose endpoint is a vertex. The intersection of two edges is either
empty or a vertex of the network. The union of the edges in E is a connected subset of Rd . For a
given edge e ∈ E , the notation ∂ e is used for the set of endpoints of e, and e∗ = e\∂ e stands for
the interior of e. Let also ue be a unit vector aligned with e. There are two possible such vectors:
if the boundary of e is made of one vertex x only, then ue will be oriented from x to the interior
of e; if the boundary of e is made of two vertices, then the choice of the orientation is arbitrary.
76
CHAPITRE 2. THE NETWORK CASE
We say that two vertices are adjacent if they are connected by an edge. For a given vertex x, we denote
by E x the set of the edges for which x is an endpoint, and Nx the cardinality of E x . We denote by G the
union of all the edges in E .
We consider infinite horizon optimal control problems which have different dynamics and running cost
in the edges. We are going to describe the assumptions on the dynamics and costs in each edge e. The
sets of controls are denoted by Ae and the system is driven by a dynamics f e and the running cost is
given by `e . Our main assumptions are as follows
[H0n ] A is a metric space (one can take A = Rm ). For e ∈ E , Ae is a non empty compact subset of A
and f e : e × Ae → R is a continuous bounded function. The sets Ae are disjoint. Moreover, there
exists L > 0 such that for any e ∈ E , x, y in e and a ∈ Ae ,
| f e (x, a) − f e ( y, a)| ≤ L|x − y|.
We will use the notation Fe (x) for the set { f e (x, a)ue , a ∈ Ae }.
[H1n ] For e ∈ E , the function `e : e×Ae → R is a continuous and bounded function. There is a modulus
of continuity ωe such that for all x, y in e and for all a ∈ Ae , |`e (x, a) − `e ( y, a)| ≤ ωe (|x − y|).
[H2n ] For e ∈ E , x ∈ e, the non empty and closed set FLe (x) ≡ {( f e (x, a)ue , `e (x, a)), a ∈ Ae } is
convex.
[H3n ] There is a real number δ > 0 such that for any e ∈ E , for all endpoints x of e,
[−δue , δue ] ⊂ Fe (x).
Let us denote by M the set:
a ∈ Ae if x ∈ e∗ , and a ∈ ∪e∈E x Ae if x ∈ V .
M = (x, a); x ∈ G ,
(2.7.1)
The set M is closed. We also define the function f on M by
∀(x, a) ∈ M ,
f (x, a) =
§
f e (x, a)ue
f e (x, a)ue
if x ∈ e∗ ,
if x ∈ V and a ∈ Ae for e ∈ E x .
The set of admissible controlled trajectories starting from the initial datum x ∈ G can be defined by
Tx =

∞
+
 ( y x , α) ∈ Lloc (R ; M ) :
y x ∈ Lip(R+ ; G ),
Z t
y x (t) = x +

f ( y x (s), α(s))ds
0


in R+ 
,
(2.7.2)
exactly as in § 2.1.1.
The cost associated to the trajectory ( y x , α) ∈ T x is
J(x; ( y x , α)) =
Z
∞
`( y x (t), α(t))e−λt d t,
0
where λ > 0 is a real number and the Lagrangian ` is defined on M by
∀(x, a) ∈ M ,
`(x, a) =
§
`e (x, a)
`e (x, a)
if x ∈ e∗ ,
if x ∈ V and a ∈ Ae for e ∈ E x .
The value function of the infinite horizon optimal control problem is
v(x) =
inf
( y x ,α)∈T x
J(x; ( y x , α)).
(2.7.3)
2.7. THE CASE OF A NETWORK
2.7.2
77
The Hamilton-Jacobi equation
For each edge e, x ∈ e∗ , let x e be the coordinate of x in the system (Oe , ue ) where Oe is an arbitrary
origin on e.
For the definition of viscosity solutions on the irregular set G , it is necessary to first define a class
of the admissible test-functions
Definition 2.7.1. A function φ : G → R is an admissible test-function if
• φ is continuous in G and C 1 in G \ V
• for any e, φ|e ∈ C 1 (e).
The set of admissible test-function is noted R(G ). If φ ∈ R(G ) and ζ ∈ R, let Dφ(x, ζue ) be defined by
dφ
dφ
Dφ(x, ζue ) = ζ d x e (x) if x ∈ e∗ , and Dφ(x, ζue ) = ζ lim y→x, y∈e∗ d x e ( y), if x is an endpoint of e.
We define the Hamiltonians H e : e × R → R by
H e (x, p) = max(−p f e (x, a) − `e (x, a)).
(2.7.4)
a∈Ae
For a vertex x ∈ V , for a given indexing of E x : E x = {e1 , . . . , eNx }, we use the notation Ai = Aei , f i = f ei ,
`i = `ei for simplicity. Let also σi be 1 if uei is oriented from x to the interior of ei and −1 in the opposite
case. The Hamiltonian H x : RNx → R is defined by
H x (p1 , . . . , pNx ) = max
i=1,...,Nx a∈Ai
max
(−pi f i (x, a) − `i (x, a)).
s.t. σi f i (x,a)≥0
(2.7.5)
We wish to define viscosity solutions of the following equations
λv(x) + H e (x, Dv(x))
=0
if x ∈ e∗ ,
(2.7.6)
λv(x) + H x (Dv(x))
=0
if x ∈ V .
(2.7.7)
Definition 2.7.2.
• An upper semi-continuous function w : G → R is a subsolution of (2.7.6)-(2.7.7)
in G if for any x ∈ G , any φ ∈ R(G ) s.t. w − φ has a local maximum point at x, then
dφ
λw(x) + H e (x, d x e (x)) ≤ 0
dφ
if x ∈ e∗ ,
dφ
λw(x) + H x ( d x 1 (x), . . . , d x N (x)) ≤ 0
x
where in the last case,
dφ
d x i (x)
if x ∈ V ,
(2.7.8)
= Dφ(x, uei (x)), for i = 1, . . . , Nx .
• A lower semi-continuous function w : G → R is a supersolution of (2.7.6)-(2.7.7) if for any x ∈ G ,
any φ ∈ R(G ) s.t. w − φ has a local minimum point at x, then
dφ
λw(x) + H e (x, d x e (x)) ≥ 0
λw(x) +
2.7.3
dφ
dφ
H x ( d x 1 (x), . . . , d x N
x
if x ∈ e∗ ,
(x)) ≥ 0
if x ∈ V .
(2.7.9)
Comparison principle
Since all the arguments used in the junction case are local, we can replicate them in the case of a
network and obtain:
Theorem 2.7.1. Assume [H0n ],[H1n ],[H2n ] and [H3n ]. Let v : G → R be a bounded viscosity subsolution
of (2.7.6)-(2.7.7), and w : G → R be a bounded viscosity supersolution of (2.7.6)-(2.7.7). Then v ≤ w in
G.
2.7.4
Existence and uniqueness
By the same arguments as in the junction case, we can prove that v is a bounded viscosity solution of
(2.7.6)-(2.7.7). From the Theorem 2.7.1, it is the unique bounded viscosity solution.
Proposition 2.7.1. Assume [H0n ],[H1n ],[H2n ] and [H3n ]. The value function v of the optimal control
problem (2.7.3) is the unique bounded viscosity solution of (2.7.6)-(2.7.7).
Remark 2.7.1. The stability results of § 2.5 for junctions can be easily generalized to networks.
78
CHAPITRE 2. THE NETWORK CASE
2.8
Appendix
2.8.1
Proof of Lemma 2.2.2
¦
Š©
S €
e
For any i ∈ {1, . . . , N }, the inclusion co FL+i (O) ∪ j6=i FL j (O) ∩ ({0} × R)
⊂ f`(O)
is proved by
explicitly constructing trajectories, see [ACCT13]. We skip this part. This leads to
¨
«
Š
[
[€
+
e
co FL (O) ∪
FL j (O) ∩ ({0} × R) ⊂ f`(O).
i
j6=i
i=1,...,N
e
We now prove the other inclusion. For any (ζ, µ) ∈ f`(O),
there exists a sequence of admissible trajectories ( yn , αn ) ∈ TO and a sequence of times t n → 0+ such that
Z tn
Z tn
1
1
f ( yn (t), αn (t))d t = ζ, and lim
`( yn (t), αn (t))d t = µ.
lim
n→∞ t
n→∞ t
n 0
n 0
• If ζ 6= 0, then there must exist an index i in {1, . . . , N } such that ζ = |ζ|ei : in this case, yn (t n ) ∈
Ji \{O}. Hence,
Z tn
Z tn
N
X
f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t
(2.8.1)
f ( yn (t), αn (t))d t =
ej
yn (t n ) =
0
with
j=1
Z
0
tn
Z0t n
f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t = 0
if j 6= i,
f i ( yn (t), αn (t))1 yn (t)∈Ji \{O} d t = | yn (t n )|.
0
These identities are a consequence of Stampacchia’s theorem: consider for example j 6= i and the
1,∞
function κ j : y 7→ | y|1 y∈J j . It is easy to check that t 7→ κ j ( yn (t)) belongs to W0 (0, t n ) and that
its weak derivative coincides almost everywhere with t 7→ f j ( yn (t), αn (t))1 yn (t)∈J j \{O} . Hence,
Z
tn
f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t = 0.
0
For j = 1, . . . , N , let T j,n be defined by
¦
©
T j,n = t ∈ [0, t n ] : yn (t) ∈ J j \{O} .
If j 6= i and T j,n > 0 then
Z t n
Z tn
1
f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t,
` j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t
T j,n
0
0
Z t n
Z tn
1
=
f j (O, αn (t))1 yn (t)∈J j \{O} d t,
` j (O, αn (t))1 yn (t)∈J j \{O} d t + o(1)
T j,n
0
0
where
o(1) is a vector tending to 0 as n → ∞. Therefore, the distance ofŠ
€ R
R tn
tn
1
T j,n e j 0 f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t, 0 ` j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t to the set FL j (O) tends
R tn
to 0. Moreover, 0 f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t = 0. Hence, the distance of
Š
€
€ Rt
R tn
n
1
T j,n e j 0 f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t, 0 ` j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t to the set FL j (O) ∩
Š
({0} × R) tends to zero as n tends to ∞.
If the set {t : yn (t) = O} has a nonzero measure, then
¨
«
Z tn
N €
Š
[
1
0,
`(O, αn (t))1{t: yn (t)=O} d t ∈ co
FL j (O) ∩ ({0} × R) .
|{t : yn (t) = O}| 0
j=1
2.8. APPENDIX
79
Finally, we know that Ti,n > 0.
tn
1
Ti,n
Z
1
=
Ti,n
Z
f i ( yn (t), αn (t))1 yn (t)∈Ji \{O} d t,
Z
0
tn
`i ( yn (t), αn (t))1 yn (t)∈Ji \{O} d t
0
tn
f i (O, αn (t))1 yn (t)∈Ji \{O} d t,
0
Z
tn
`i (O, αn (t))1 yn (t)∈Ji \{O} d t + o(1)
0
so the
distance of
€ R
Š
R tn
tn
1
e
f
(
y
(t),
α
(t))1
d
t,
`
(
y
(t),
α
(t))1
d
t
to the set FL+i (O) tends
i
i
n
n
y
(t)∈J
\{O}
i
n
n
y
(t)∈J
\{O}
Ti,n
n
i
n
i
0
0
to zero as n tends to ∞.
Combining
all the observations above, we see that
Š the distance of
€ Rt
R tn
n
1
1
t n 0 f ( yn (t), αn (t))d t, t n 0 `( yn (t), αn (t))d t
¦
Š©
S €
to co FL+i (O) ∪ j6=i FL j (O) ∩ ({0} × R) tends to 0 as n → ∞.
¦
Š©
S €
Therefore (ζ, µ) ∈ co FL+i (O) ∪ j6=i FL j (O) ∩ ({0} × R) .
• If ζ = 0, either there exists i such that yn (t n ) ∈ Ji \{O} or yn (t n ) = O:
• If yn (t n ) ∈¦ Ji \{O}, then we
can make exactlyŠ©
the same argument as above and conclude that
S €
(ζ, µ) ∈ co FL+i (O) ∪ j6=i FL j (O) ∩ ({0} × R) . Since ζ = 0, we have in fact that (ζ, µ) ∈
Š
SN €
co j=1 FL j (O) ∩ ({0} × R) .
Z tn
f j ( yn (t), αn (t))1 yn (t)∈J j \{O} d t = 0 for all j = 1, . . . , N . We can
• If yn (t n ) = O, we have that
0
repeat the ¦
argument
above, and obtain
Š© that
SN €
(ζ, µ) ∈ co
.
j=1 FL j (O) ∩ ({0} × R)
80
CHAPITRE 2. THE NETWORK CASE
CHAPITRE 3
HAMILTON-JACOBI EQUATIONS FOR
OPTIMAL CONTROL ON
MULTIDIMENSIONAL JUNCTIONS
Résumé
On considère des problèmes de contrôle optimal pour lesquels l’état est
contraint à rester sur une jonction, c’est-à-dire une union finie de demihyperplans partageant le même bord. On introduit une notion de solution
de viscosité pour des équations de Hamilton-Jacobi sur la jonction et on propose une preuve d’un principe de comparaison reposant en partie sur des
arguments issus de la théorie du contrôle optimal.
Abstract
We consider continuous-state and continuous-time control problems where
the admissible trajectories of the system are constrained to remain on a union
of half-planes which share a common straight line. This set will be named a
junction. We define a notion of constrained viscosity solution of HamiltonJacobi equations on the junction and we propose a comparison principle
whose proof is based on arguments from the optimal control theory.
3.1
Introduction
We are interested in optimal control problems whose trajectories are constrained to remain on a multidimensional junction. We define a junction in Rd , d ≥ 2, as a union of half-hyperplanes sharing an affine
space of dimension d − 2, see Figure 3.1 for a junction in R3 . For simplicity, we shall limit ourselves to
junctions in R3 , although all what follows can be generalized for d ≥ 3. We shall name interface the
straight line Γ shared by the half-planes.
P2
P3
Γ
e3
e2 e
0
e1
P1
eN
e4
P4
PN
Figure 3.1: The junction S in R3
81
82
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
The junctions in Rd are particular ramified sets, which we define as closed and connected subsets
of Rd obtained as the union of embedded manifolds with dimension strictly smaller than d. Figure 3.2
below supplies examples of ramified sets.
•
•
•
•
•
•
(a)
(b)
(c)
Figure 3.2: Examples of ramified sets
While there is a wide literature on control problems with state constrained in closures of open sets
( [Son86a, Son86b], [CDL90], [IK96]), the interest on problems with state constrained in closed sets
with empty interior is more recent. The results of Frankowska and Plaskacz [FP00b, FP00a] do apply
to some closed sets with empty interior, but not to ramified sets except in very particular cases.
The case of Hamilton-Jacobi equations on networks, see Figure 3.2 (a), is now well understood. The
first work seems to be the thesis of Schieborn in 2006, [Sch06]. It was focused on eikonal equations.
These results were improved later in [SC13]. The notion of viscosity solutions in these works are restricted to eikonal equations and cannot be used for more general control problems. The first two
articles on optimal control problems whose dynamics are constrained to a network were published
in 2013: in [ACCT13] Achdou, Camilli, Cutrì and Tchou proposed a Hamilton-Jacobi equation and
a definition of viscosity solution. Independently, in [IMZ13] Imbert, Monneau and Zidani proposed
a Hamilton-Jacobi approach to junction problems and traffic flows and gave an equivalent notion of
viscosity solution. Both [ACCT13] and [IMZ13] contain the first comparison and uniqueness results
for Hamilton-Jacobi equations on networks, but these results needed rather strong assumptions on the
Hamiltonians behaviour. A general comparison result has finally been obtained in the recent paper
by Imbert-Monneau [IM14b]. In the latter, the Hamiltonians in the edges are completely independent
from each other; the main assumption is that the Hamiltonian in each edge, say H i (x, p) for the edge
indexed i, is coercive and bimonotone, i.e. non increasing (resp. non decreasing) for p smaller (resp.
larger) than a given threshold pi0 (x). Of course, convex Hamiltonian coming from optimal control theory are bimonotone. Moreover, in [IM14b], the authors consider more general transmission conditions
than in [ACCT13, IMZ13], allowing an additional running cost at the junctions. Soon after, Y. Achdou,
S. Oudet and N. Tchou, [AOT15b], proposed a different proof of a general comparison result in the case
of control problems. In [AOT15b], the dynamics and running costs be different on each edges, and the
Hamiltonians in the edges are a priori completely independent from each other, as in [IM14b]. Whereas
the proof of the comparison result in [IM14b] it is only based on arguments from the theory of partial
differential equations, the proof in [AOT15b] is based on arguments from the theory of control which
were first introduced by G. Barles, A. Briani and E. Chasseigne in [BBC13,BBC14]. In the latter articles,
the authors study control problems in RN whose dynamics and running costs may be discontinuous
across an hyperplane. The problems studied in [BBC13, BBC14] and in [AOT15b] have the common
point that the data are discontinuous across a low dimensional subregion.
There is even less literature on Hamilton-Jacobi equations posed on more general ramified spaces. In
their recent article [CSM13], F. Camilli, D. Schieborn and C. Marchi deal with eikonal equations, generalize the special notion of viscosity solutions proposed in [Sch06, SC13], and prove existence and
uniqueness theorems. The work by Y. Giga, N. Hamamuki and A. Nakayasu, see [GHN14], is devoted
to eikonal equations in general metric spaces, and their results apply to ramified spaces. The difficulty
with general metric space (χ, d) is that the gradient Du of a function u : χ → R is not well-defined in
general. Yet, eikonal equation can be studied since a definition for the modulus of the gradient can be
given. More general results in geodesic metric spaces have been recently given by L. Ambrosio and J.
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
83
Feng, see [AF14], see also the recent paper of [Nak14], who considered evolutionary Hamilton-Jacobi
equations of the form u t + H(x, |Du|) = 0 in a metric space. For optimal control problems on ramified
sets, we can mention the recent article by C. Hermosilla and H. Zidani [HZ15], in which they study infinite horizon problems whose trajectories are constrained to remain in a set with a stratified structure.
The authors obtain existence and uniqueness results with weak controllability assumptions, but they
assume that the dynamics is continuous.
The present work is a continuation of [AOT15b] (which was focused on networks), but since the interface Γ is now a straight line instead of a point, the trajectories that stay on Γ have a richer structure
than in [AOT15b]. We will have to introduce a tangential Hamiltonian HΓT to take these admissible
trajectories into account. Different controllability assumptions can be made
1. strong controllability in a neighborhood of Γ
2. a weaker controllability assumption in a neighborhood of Γ , namely normal controllability to Γ ,
e 3 ] in § 3.2.1.2.
see [H
As in [AOT15b], the proof of the comparison results will be inspired by the arguments contained
in [BBC13, BBC14].
In § 3.2, we discuss the case when strong controllability is assumed in a neighborhood of Γ . We
propose a Bellman equation and a notion of viscosity solutions, prove that the value function is indeed a
continuous viscosity solution of this equation, and give a comparison result. In § 3.3, the same program
is carried out when only normal controllability holds in a neighborhood of Γ . As in [BBC13], we first
prove that the value function is a discontinuous viscosity solution of the Bellman equation. We then
prove a comparison result. The latter implies the continuity of the value function. Finally, in § 3.4, we
extend the results by assuming that in addition to the dynamics and costs related to the hyperplanes,
there is a pair of tangential dynamics and tangential running cost defined on Γ .
Although we will not discuss it, the results obtained below can be generalized to ramified sets for which
the interfaces are non intersecting manifolds of dimension d − 2, see for example Figure 3.2 (b). On
the contrary, it is not obvious to apply them to the ramified sets for which interfaces of dimension d − 2
cross each other, see Figure 3.2 (c). This topic will hopefully be discussed in a forecoming work.
3.2
First case : full controllability near the interface
3.2.1
Setting of the problem and basic assumptions
3.2.1.1
The geometry
We are going to study optimal control problems in Rd , d = 3, with constraints on the state of the
system. The state is constrained to lie in the union S of N half-planes, N > 1. Let (ei )i=0,...,N , be some
respectively distinct unit vectors in Rd such that ei .e0 = 0 for any i ∈ {1, . . . , N }, where for x, y ∈ Rd ,
x. y denotes the usual scalar product of the Euclidean space Rd . The notation |.| will be used for the
usual Euclidean norm in Rd . For i = 1, . . . , N , Pi is the closed half-plane Re0 × R+ ei . We denote by Γ
the straight line Re0 . The half-planes Pi are glued at the straight line Γ to form the set S , see Figure
3.1:
N
[
S =
Pi .
i=1
For any x ∈ S , we denote by Tx (S ) ⊂ R the set of the tangent directions to S , i.e. Tx (S ) =
Re0 × Rei , for any x ∈ Pi \Γ and Tx (S ) = ∪Ni=1 (Re0 × Rei ) for any x ∈ Γ .
If x ∈ S \ Γ , ∃! i ∈ {1, . . . , N }, ∃! x 0 ∈ R and ∃! x i ∈ R+ \ {0} such that
d
x = x 0 e0 + x i ei .
(3.2.1)
We will use often this decomposition. Sometime, it will be convenient to extend this decomposition to
the whole set S , by writing x = x 0 e0 + 0ei for any i ∈ {1, . . . , N } if x ∈ Γ . When we will not want to
specify in which half-plan Pi is x belonging to S , we will use the notation
x = x 0 e0 + x 0 ,
(3.2.2)
84
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
where x 0 denotes x i ei if x = x 0 e0 + x i ei .
The geodesic distance d(x, y) between two points x, y of S is
§
|x − y| if x, y belong to the same half-plane Pi
d(x, y) =
minz∈Γ {|x − z| + |z − y|} if x, y belong to different half-planes Pi and P j .
(3.2.3)
More classically, if x ∈ Rd and C is a closed subset of Rd , dist(x, C) will denote
d ist(x, C) = inf{|x − z| : z ∈ C},
the distance between x and C. The notation B(Γ , r) will be used to denote the set {x ∈ Rd : dist(x, Γ ) <
r}.
3.2.1.2
The optimal control problem
We consider infinite-horizon optimal control problems which have different dynamics and running costs
in the half-planes. We are going to describe the assumptions on the dynamics and costs in each halfplane Pi : the sets of controls are denoted by Ai , the system is driven by the dynamics f i and the running
costs are given by `i . Our main assumptions are as follows
[H0] A is a metric space (one can take A = Rm ). For i = 1, . . . , N , Ai is a non empty compact subset
of A and f i : Pi × Ai → Re0 × Rei is a continuous bounded function. The sets Ai are disjoint.
Moreover, there exists L f > 0 such that for any i, x, y ∈ Pi and a ∈ Ai ,
| f i (x, a) − f i ( y, a)| ≤ L f |x − y|.
We note M f the minimal constant such that for any x ∈ S , i ∈ {1, . . . , N } and a ∈ Ai ,
| f i (x, a)| ≤ M f .
We will use also the notation Fi (x) for the set { f i (x, a), a ∈ Ai }.
[H1] For i = 1, . . . , N , the function `i : Pi × Ai → R is a continuous and bounded function. There is a
modulus of continuity ω` such that for all i ∈ {1, . . . , N }, x, y ∈ Pi and a ∈ Ai ,
|`i (x, a) − `i ( y, a)| ≤ ω` (|x − y|).
We denote M` the minimal constant such that for any i ∈ {1, . . . , N }, x ∈ Pi and a ∈ Ai ,
|`i (x, a)| ≤ M` .
[H2] For i = 1, . . . , N , x ∈ Pi , the non empty and closed set
FLi (x) = {( f i (x, a), `i (x, a)), a ∈ Ai }
is convex.
[H3] There is a real number δ > 0 such that for any i = 1, . . . , N and for all x ∈ Γ ,
B(0, δ) ∩ (Re0 × Rei ) ⊂ Fi (x).
In § 3.3 below, we will weaken assumption [H3] and use only the assumption on normal controllability
e 3] There is a real number δ > 0 such that for any i = 1, . . . , N and for all x ∈ Γ ,
[H
[−δ, δ] ⊂ { f i (x, a).ei : a ∈ Ai } .
Remark 3.2.1. In [H0] the assumption that the sets Ai are disjoint is not restrictive: it is made only for
simplifying the proof of Theorem 3.2.2 below. The assumption [H2] is not essential : it is made in order to
avoid the use of relaxed controls.
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
85
Thanks to the Filippov implicit function lemma, see [MW67], we obtain:
Theorem 3.2.1. Let I be an interval of R and γ : I → Rd × Rd be a measurable function. Let K be a closed
subset of Rd × A and Ψ : K → Rd × Rd be continuous. Assume that γ(I) ⊂ Ψ(K), then there is a measurable
function Φ : I → K with
Ψ ◦ Φ(t) = γ(t) for a.a. t ∈ I.
Let M denote the set:
M = (x, a); x ∈ S ,
a ∈ Ai if x ∈ Pi \Γ , and a ∈ ∪Ni=1 Ai if x ∈ Γ .
The set M is closed. We also define the function f on M by
§
f i (x, a)
∀(x, a) ∈ M ,
f (x, a) =
f i (x, a)
if x ∈ Pi \Γ ,
if x ∈ Γ and a ∈ Ai .
(3.2.4)
(3.2.5)
Remark 3.2.2. The function f is well defined on M , because the sets Ai are disjoint, and is continuous on
M.
Let F̃ (x) be defined by
§
Fi (x)
F̃ (x) =
∪Ni=1 Fi (x)
if x belongs to the open half-plane Pi \Γ
if x ∈ Γ .
For x ∈ S , the set of admissible trajectories starting from x is
ẏ (t) ∈ F̃ ( y (t)),
x
+
Yx = y x ∈ Lip(R ; S ) : x
y x (0) = x,
for a.a. t > 0,
,
(3.2.6)
where Lip(R+ ; S ) is the set of the Lipschitz continuous functions from R+ to S .
As in [BBC13] and [AOT15b] the following statement holds :
Theorem 3.2.2. Assume [H0], [H1], [H2] and [H3]. Then
1. For any x ∈ S , Yx is not empty.
2. For any x ∈ S , for each trajectory y x in Yx , there exists a measurable function Φ : [0, +∞) → M ,
Φ(t) = (ϕ1 (t), ϕ2 (t)) with
( y x (t), ẏ x (t)) = (ϕ1 (t), f (ϕ1 (t), ϕ2 (t))),
for a.a. t,
which means in particular that y x is a continuous representation of ϕ1 .
3. Almost everywhere on {t : y x (t) ∈ Γ }, f ( y x (t), ϕ2 (t)) ∈ Re0 .
We introduce the set of admissible controlled trajectories starting from the initial datum x :


y x ∈ Lip(R+ ; S ),


Z
t
∞
T x = ( y x , α) ∈ LLoc
(R+ ; M ) : .
(3.2.7)
f ( y x (s), α(s))ds in R+ 

y x (t) = x +
0
Remark 3.2.3. If two different half-planes are parallel to each other, say the half-planes P1 and P2 , many
other assumptions can be made on the dynamics and costs:
• a trivial case in which the assumptions [H1]-[H3] are satisfied is when the dynamics and costs are
continuous at the origin, i.e. A1 = A2 ; f1 and f2 are respectively the restrictions to P1 × A1 and
P2 × A2 of a continuous and bounded function f1,2 defined in Re0 × Re1 × A1 , which is Lipschitz
continuous with respect to the first variable; `1 and `2 are respectively the restrictions to P1 × A1 and
P2 × A2 of a continuous and bounded function `1,2 defined in Re0 × Re1 × A1 .
• In this particular geometrical setting, one can allow some mixing (relaxation) at the vertex with
several possible rules: More precisely, in [BBC13, BBC14], Barles et al. introduce several kinds of
trajectories which stay at the interface: the regular trajectories are obtained by mixing outgoing
dynamics from P1 and P2 , whereas singular trajectories are obtained by mixing strictly ingoing
dynamics from P1 and P2 . Two different value functions are obtained whether singular mixing is
permitted or not.
86
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
The cost functional
The cost associated to the trajectory ( y x , α) ∈ T x is
∞
Z
J(x; ( y x , α)) =
`( y x (t), α(t))e−λt d t,
(3.2.8)
0
where λ > 0 is a real number and the Lagrangian ` is defined on M by
`(x, a) =
∀(x, a) ∈ M ,
The value function
§
`i (x, a)
`i (x, a)
if x ∈ Pi \Γ ,
if x ∈ Γ and a ∈ Ai .
(3.2.9)
The value function of the infinite horizon optimal control problem is
v(x) =
inf
( y x ,α)∈T x
J(x; ( y x , α)).
(3.2.10)
Proposition 3.2.1. Assume [H0] and [H1]. We have the dynamic programming principle:
∀t ≥ 0,
v(x) =
t
Z
( y x ,α)∈T x
`( y x (s), α(s))e−λs ds + e−λt v( y x (t)) .
inf
(3.2.11)
0
Proposition 3.2.2. Assume [H0], [H1], [H2] and [H3]. Then the value function v is bounded and continuous on S .
Both propositions above are classical. See appendices § 3.5.1 and § 3.5.2 for the proofs.
3.2.2
The Hamilton-Jacobi equation
3.2.2.1
Test-functions
To define viscosity solutions on the irregular set S , it is necessary to first define a class of admissible
test-functions
Definition 3.2.1. A function ϕ : S → R is an admissible test-function if
• ϕ is continuous in S
• for any j, j = 1, . . . , N , ϕ|P j ∈ C 1 (P j ).
The set of admissible test-functions is denoted by R(S ). If ϕ ∈ R(S ), x ∈ S and ζ ∈ Tx (S ), let Dϕ(x, ζ)
be defined by
Dϕ(x, ζ) =
§
D(ϕ|Pi )(x).ζ
D(ϕ|Pi )(x).ζ
if
if
x ∈ Pi \Γ ,
x ∈ Γ and ζ ∈ Re0 × Rei ,
where
for u, v ∈ Re0 × Rei , u.v denotes the usual Euclidean scalar product in Re0 × Rei and for x ∈ Γ ,
D ϕ|Pi (x).ξ is defined by
D ϕ|Pi (x).ξ =
lim
y→x, y∈Pi \Γ
D ϕ|Pi ( y).ξ.
(3.2.12)
If ϕ ∈ R(S ), x ∈ Γ and ξ ∈ Re0 , we will use also the notations D(ϕ|Γ )(x).ξ for the differential of ϕ|Γ at
the point x evaluated in ξ.
Other notations which will be useful are the following: for j ∈ {1, . . . , N }, x ∈ P j and φ ∈ R(S )
(
€
Š
∂ x j φ|P j (x) =
limh→0
lim h>0
h→0
and for x ∈ Γ
φ(x+he j )−φ(x)
h
φ(x+he j )−φ(x)
h
if x ∈ Pi \ Γ ,
if x ∈ Γ ,
φ(x + he0 ) − φ(x)
.
h→0
h
∂ x 0 φ(x) = lim
(3.2.13)
(3.2.14)
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
87
Remark 3.2.4. If ϕ ∈ R(S ), x ∈ Γ and ξ ∈ R, then for all i ∈ {1, . . . , N },
D(ϕ|Pi )(x).ξe0 = D(ϕ|Γ )(x).ξe0 = ∂ x 0 φ(x)ξ.
Particularly, the tangential component of D(ϕ|Pi )(x) is independent of i ∈ {1, . . . , N }.
Property 3.2.1. If ϕ = g ◦ ψ with g ∈ C 1 (R) and ψ ∈ R(S ), then ϕ ∈ R(S ) and for any x ∈ S ,
ζ ∈ Tx (S )
Dϕ(x, ζ) = g 0 (ψ(x))Dψ(x, ζ).
3.2.2.2
Vector fields
For i = 1, . . . , N and x in Γ , we denote by Fi+ (x) and FL+i (x) the sets
FL+i (x) = FLi (x) ∩ (Re0 × R+ ei ) × R ,
Fi+ (x) = Fi (x) ∩ Re0 × R+ ei ,
which are non empty thanks to assumption [H3]. Note that 0 ∈ ∩Ni=1 Fi (x). From assumption [H2],
these sets are compact and convex. For x ∈ S , the sets F (x) and FL(x) are defined by
F (x) =
§
Fi (x)
∪i=1,...,N Fi+ (x)
FL(x) =
§
FLi (x)
∪i=1,...,N FL+i (x)
and
3.2.2.3
if x belongs to Pi \Γ
if x ∈ Γ
if x belongs to Pi \Γ
if x ∈ Γ .
Definition of viscosity solutions
We now introduce the definition of a viscosity solution of
λu(x) +
sup
{−Du(x, ζ) − ξ} = 0
(ζ,ξ)∈FL(x)
in S .
(3.2.15)
Definition 3.2.2.
• An upper semi-continuous function u : S → R is a subsolution of (3.2.15) in S
if for any x ∈ S , any ϕ ∈ R(S ) s.t. u − ϕ has a local maximum point at x, then
λu(x) +
sup
{−Dϕ(x, ζ) − ξ} ≤ 0.
(3.2.16)
(ζ,ξ)∈FL(x)
• A lower semi-continuous function u : S → R is a supersolution of (3.2.15) if for any x ∈ S , any
ϕ ∈ R(S ) s.t. u − ϕ has a local minimum point at x, then
λu(x) +
sup
{−Dϕ(x, ζ) − ξ} ≥ 0.
(3.2.17)
(ζ,ξ)∈FL(x)
• A continuous function u : S → R is a viscosity solution of (3.2.15) in S if it is both a viscosity
subsolution and a viscosity supersolution of (3.2.15) in S .
Remark 3.2.5. At x ∈ Pi \Γ , the notion of sub, respectively super-solution in Definition 3.2.2 is equivalent
to the standard definition of viscosity sub, respectively super-solution of
λu(x) + sup{− f i (x, a) · Du(x) − `i (x, a)} = 0.
a∈Ai
3.2.2.4
Hamiltonians
We define the Hamiltonian H i : Pi × (Re0 × Rei ) → R by
H i (x, p) = max(−p. f i (x, a) − `i (x, a)),
a∈Ai
(3.2.18)
88
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
and the Hamiltonian HΓ : Γ ×
Q
i=1,...,N (Re0
× Rei ) → R by
HΓ (x, p1 , . . . , pN ) = max H i+ (x, pi ),
i=1,...,N
(3.2.19)
where the Hamiltonian H i+ : Pi × (Re0 × Rei ) → R is defined by
H i+ (x, p) =
a∈Ai
max
(−p. f i (x, a) − `i (x, a)).
s.t. f i (x,a).ei ≥0
(3.2.20)
We also define what may be called the tangential Hamiltonian at Γ , HΓT : Γ × Re0 → R, by
HΓT (x, p) = max HΓT,i (x, p),
i=1,...,N
(3.2.21)
where the Hamiltonian HΓT,i : Γ × Re0 → R is defined by
HΓT,i (x, p) =
a∈Ai
max
(− f i (x, a).p − `i (x, a)).
s.t. f i (x,a).ei =0
(3.2.22)
The following definitions are equivalent to Definition 3.2.2:
Definition 3.2.3.
• An upper semi-continuous function u : S → R is a subsolution of (3.2.15) in S
if for any x ∈ S , any ϕ ∈ R(S ) s.t. u − ϕ has a local maximum point at x, then
λu(x) + H i (x, D ϕ|Pi (x)) ≤ 0
if x ∈ Pi \Γ ,
(3.2.23)
if x ∈ Γ .
λu(x) + HΓ x, D ϕ|P1 (x), . . . , D ϕ|PN (x) ≤ 0
• A lower semi-continuous function u : S → R is a supersolution of (3.2.15) if for any x ∈ S , any
ϕ ∈ R(S ) s.t. u − ϕ has a local minimum point at x, then
λu(x) + H i (x, D ϕ|Pi (x)) ≥ 0
if x ∈ Pi \Γ ,
(3.2.24)
if x ∈ Γ .
λu(x) + HΓ x, D ϕ|P1 (x), . . . , D ϕ|PN (x) ≥ 0
The Hamiltonian H i are continuous with respect to x ∈ Pi , convex with respect to p. Moreover, if
x belongs to Γ , the function p 7→ H i (x, p) is coercive, i.e. lim|p|→+∞ H i (x, p) = +∞ from the controllability assumption [H3].
The following lemma is the counterpart of Lemma 2.1 in [AOT15b].
Lemma 3.2.1. Assume [H0], [H1], [H2] and [H3]. Take i ∈ {1, . . . , N }, x ∈ Γ and p ∈ Re0 × Rei . Let
ϕi,x,p : R → R be the function defined by ϕi,x,p (δ) = H i (x, p + δei ). We denote by ∆i,x,p the set
∆i,x,p = {δ ∈ R s.t. ϕi,x,p (δ) = min(ϕi,x,p (d))}.
d∈R
(3.2.25)
1. The set ∆i,x,p is not empty.
2. δ ∈ ∆i,x,p if and only if there exists a∗ ∈ Ai such that f i (x, a∗ ).ei = 0 and
ϕi,x,p (δ) = − f i (x, a∗ ).p − `i (x, a∗ ).
3. For any x ∈ Γ , p = p0 e0 + pi ei , with p0 , pi ∈ R and δ ∈ ∆i,x,p we have
H i (x, p + δei ) = H i+ (x, p + δei ) = HΓT,i (x, p0 e0 ).
(3.2.26)
4. For any x ∈ Γ , p = p0 e0 + pi ei , with p0 , pi ∈ R and δ ≥ min{d : d ∈ ∆i,x,p } we have
H i+ (x, p + δei ) = HΓT,i (x, p0 e0 ).
(3.2.27)
Proof. Point 1 is easy, because the Hamiltonian H i is continuous and coercive with respect to p. The
function ϕi,x,p reaches its minimum at δ if and only if 0 ∈ ∂ ϕi,x,p (δ). The subdifferential of ϕi,x,p at δ
is characterized by
∂ ϕi,x,p (δ) = co{− f i (x, a).ei ; a ∈ Ai s.t. ϕi,x,p (δ) = − f i (x, a).(p + δei ) − `i (x, a)},
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
89
see [Val69]. But from [H2],
{( f i (x, a), `i (x, a)); a ∈ Ai s.t. ϕi,x,p (δ) = − f i (x, a).(p + δei ) − `i (x, a)}
is compact and convex. Hence,
∂ ϕi,x,p (δ) = {− f i (x, a).ei ; a ∈ Ai s.t. ϕi,x,p (δ) = − f i (x, a).(p + δei ) − `i (x, a)}.
Therefore, 0 ∈ ∂ ϕi,x,p (δ) if and only if there exists a∗ ∈ Ai such that f i (x, a∗ ).ei = 0 and ϕi,x,p (δ) =
− f i (x, a∗ ).(p + δei ) − `i (x, a∗ ). We have proved point 2.
Points 3 is a direct consequence of point 2. Point 4 is a consequence of point 3 and of the decreasing
character of the function d 7−→ H i+ (x, p + dei ).
e 3]. Indeed, we actually
Remark 3.2.6. The conclusions of Lemma 3.2.1 hold if we replace [H3] with [H
just need that the Hamiltonian H i be coercive with respect to pi , where pi = p.ei .
3.2.2.5
Existence
Theorem 3.2.3. Assume [H0], [H1], [H2] and [H3]. The value function v defined in (3.2.10) is a bounded
viscosity solution of (3.2.15) in S .
The proof of Theorem 3.2.3 is made in several steps: the first step consists of proving that the value
function is a viscosity solution of a Hamilton-Jacobi equation with a more general definition of the
Hamiltonian: for that, we introduce larger relaxed vector fields: for x ∈ S ,


t n → 0+ and
∃( y x,n , αn )n∈N ,


Z
t
n
1
f˜(x) = η ∈ Rd : ( y x,n , αn ) ∈ T x , s.t. f ( y x,n (t), αn (t))d t = η 
lim

n→∞
∃(t n )n∈N
tn 0
and
e
f`(x)
=

t n → 0+ ,

Z tn



1
∃( y x,n , αn )n∈N ,

lim
f ( y x,n (t), αn (t))d t = η,
(η, µ) ∈ Rd × R : ( y x,n , αn ) ∈ T x , s.t. n→∞ t n 0
Z tn


∃(t n )n∈N

lim 1

`( y x,n (t), αn (t))d t = µ

n→∞
tn 0






;





where T x is the set of admissible controlled trajectories starting from the point x which was introduced
in (3.2.7).
Proposition 3.2.3. Assume [H0], [H1], [H2] and [H3]. The value function v defined in (3.2.10) is a
viscosity solution of
λu(x) + sup {−Du(x, ζ) − ξ} = 0 in S ,
(3.2.28)
e
(ζ,ξ)∈ f`(x)
where the definition of viscosity solution is exactly the same as in Definition 3.2.2, replacing FL(x) with
e
f`(x).
Proof. See appendix § 3.5.3.
The second step consists of proving that for any ϕ ∈ R(S ) and x ∈ S , sup(ζ,ξ)∈FL(x) {−Dϕ(x, ζ)−ξ}
and sup(ζ,ξ)∈ f`(x)
{−Dϕ(x, ζ) − ξ} are equal. This is a consequence of the following lemma.
e
Lemma 3.2.2. For any x ∈ S ,
e
f`(x)
=
e
f`(x) =
FL(x)
S
¦
Š©
S €
+
i=1,...,N co FLi (x) ∪
j6=i FL j (x) ∩ (Re0 × R)
Proof. The proof being a bit long, we postpone it to the appendix § 3.5.4.
if x ∈ S \Γ ,
if x ∈ Γ .
90
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Lemma 3.2.3. Assume [H0], [H1], [H2] and [H3]. For any function ϕ ∈ R(S ) and x ∈ S ,
{−Dϕ(x, ζ) − ξ} =
sup
e
(ζ,ξ)∈ f`(x)
max {−Dϕ(x, ζ) − ξ}.
(3.2.29)
(ζ,ξ)∈FL(x)
e
Proof. For x ∈ S \Γ there is nothing to prove because FL(x) = f`(x).
If x ∈ Γ we can prove that
e
FL(x) ⊂ f`(x) for any x ∈ Γ , in the same way as in [ACCT13]. Hence
max {−Dϕ(x, ζ) − ξ} ≤
(ζ,ξ)∈FL(x)
{−Dϕ(x, ζ) − ξ}.
sup
e
(ζ,ξ)∈ f`(x)
From the piecewise linearity of the function (ζ, µ) 7→ −Dϕ(x, ζ) − µ, we infer that
(ζ,µ)∈co FL+
i (x)∪
= max
sup€
S
j6=i
max+
Š(−Dϕ(x, ζ) − µ)
FL j (x)∩(Re0 ×R)
(−Dϕ(x, ζ) − µ), max
j6=i
(ζ,µ)∈FLi (x)
max
(−Dϕ(x, ζ) − µ)
(ζ,µ)∈FL j (x)∩(Re0 ×R)
≤ max j=1,...,N max(ζ,µ)∈FL+j (x) (−Dϕ(x, ζ) − µ) = max(ζ,ξ)∈FL(x) {−Dϕ(x, ζ) − ξ}.
We conclude by using Lemma 3.2.2.
3.2.3
Properties of viscosity sub and supersolutions
In this part, we study sub and supersolutions of (3.2.15), transposing ideas coming from Barles-BrianiChasseigne [BBC13, BBC14] to the present context.
Property 3.2.2. Assume [H0] and [H3]. Then, there exists R > 0 a positive real number such that for all
i = 1, . . . , N and x ∈ B(Γ , R) ∩ Pi
B(x,
δ
) ∩ (Re0 × Rei ) ⊂ Fi (x).
2
(3.2.30)
Remark 3.2.7. This property means that the controlabillity assumption [H3], which focuses on Γ , holds
in a neighborhood of Γ thanks to the continuity properties of the functions f i , i ∈ {1, . . . , N }.
Lemma 3.2.4. Assume [H0], [H1], [H2] and [H3]. Let R > 0 be as in (3.2.30). For any bounded viscosity
subsolution u of (3.2.15), there exists a constant C ∗ > 0 such that u is a viscosity subsolution of
|Du(x)| ≤ C ∗
in B(Γ , R) ∩ S ,
(3.2.31)
i.e. for any x ∈ B(Γ , R) ∩ S and φ ∈ R(S ) such that u − φ has a local maximum point at x,
|D(φ|Pi )(x)| ≤ C ∗
max |∂ x 0 φ(x)| + ∂ x i (φ|Pi )(x) − ≤ C ∗
if x ∈ (B(Γ , R) ∩ Pi )\Γ ,
(3.2.32)
if x ∈ Γ ,
(3.2.33)
i=1,...,N
€
Š
where ∂ x j φ|P j (x) and ∂ x 0 φ(x) are defined in (3.2.13) and (3.2.14) and [·]− denote the negative part
function, i.e. for x ∈ R, [x]− = max{0, −x}.
Proof. Let Mu (resp M` ) be an upper bound on |u| (resp. ` j for all j = 1, . . . , N ). The viscosity inequality
(3.2.23) yields that
H i (x, D(φ|Pi )(x))
HΓ (x, D(φ|P1 )(x), . . . , D(φ|PN )(x))
≤ λMu
≤ λMu
if x ∈ (B(Γ , R) ∩ Pi )\Γ ,
(3.2.34)
if x ∈ Γ .
(3.2.35)
From the controllability in B(Γ , R) ∩ Pi , we see that H i is coercive with respect to its second argument
uniformly in x ∈ B(Γ , R) ∩ Pi . More precisely we have that H i (x, p) ≥ δ2 |p| − M` .
λM +M
Thus, if x ∈ (B(Γ , R) ∩ Pi )\Γ , from (3.2.34), we have |D(φ|Pi )(x)| ≤ C ∗ with C ∗ = 2 uδ ` .
If x ∈ Γ , considering the controls a0+ , a0− , ai+ and ai0 ∈ Ai such that f i (x, a0± ) = ±δe0 , f i (x, ai+ ) = δei and
f i (x, ai0 ) = 0 we can prove that H i+ (x, p0 e0 + pi ei ) ≥ δ2 (|p0 | + [pi ]− ) − M` . Then, the viscosity inequality
(3.2.35) yields (3.2.33) with the same constant C ? = 2
λMu +M`
.
δ
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
91
The following lemma gives us an explicit expression for the geodesic distance which will be convenient in future calculations.
Lemma 3.2.5. Let x = x 0 e0 + x i ei ∈ Pi and y = y0 e0 + y j e j ∈ P j . Then,
( 1
(x 0 − y0 )2 + (x i − yi )2 2 if i = j,
d(x, y) =
1
(x 0 − y0 )2 + (x i + y j )2 2 if i 6= j.
(3.2.36)
Lemma 3.2.6. Assume [H0], [H1], [H2] and [H3]. Any bounded viscosity subsolution u of (3.2.15) is
Lipschitz continuous in a neighborhood of Γ , i.e there exists some strictly positive number r such that u is
Lipschiz continuous on B(Γ , r) ∩ S , where B(Γ , r) denotes the set { y ∈ Rd : dist( y, Γ ) < r}.
Proof. We adapt the proof of H.Ishii, see [Ish13].
Take R as in (3.2.30), fix z ∈ B(Γ , R) ∩ S and set r = (R − dist(z, Γ ))/4. Fix any y ∈ S such that
d( y, z) < r. It can be checked that for any x ∈ S , if d(x, y) < 3r then dist(x, Γ ) < R. Choose a function
f ∈ C 1 ([0, 3r)) such that f (t) = t in [0, 2r], f 0 (t) ≥ 1 for all t ∈ [0, 3r) and lim t→3r f (t) = +∞. Fix
any " > 0. We are going to show that
u(x) ≤ u( y) + (C ∗ + ") f (d(x, y)),
∀x ∈ S such that d(x, y) < 3r,
(3.2.37)
where C ∗ is the constant in Lemma 3.2.4. Let us proceed by contradiction. Assume that (3.2.37)
is not true. According to the properties of f , the function x 7→ u(x) − u( y) − (C ∗ + ") f (d(x, y))
admits a maximum ξ ∈ B( y, 3r) ∩ S . However, since we assumed that (3.2.37) is not true, necessarily
ξ 6= y. Consequently, the function ψ : S → R, x 7−→ (C ? + ") f (d(x, y)) is an appropriate test
function in a neighborhood of ξ which can be used as test function in the viscosity inequality (3.2.31),
satisfied by u from Lemma 3.2.4. For the calculations, we need to distinguish several cases. Assume
that y = y0 e0 + yi ei ∈ Pi .
1. If ξ = ξ0 e0 + ξi ei ∈ Pi \ Γ : i.e. ξ and y belong to the same half-plane Pi . Then, from (3.2.36)
ξ− y
in Lemma 3.2.5, we have D(ψ|Pi )(ξ) = (C ? + ") f 0 (d(ξ, y)) d(ξ, y) and (3.2.32) in Lemma 3.2.4
gives us
|ξ − y|
(C ∗ + ") f 0 (d(ξ, y))
≤ C ∗.
(3.2.38)
d(ξ, y)
Since
|ξ− y|
d(ξ, y)
= 1 and f 0 (t) ≥ 1 for all t ∈ [0, 3r), (3.2.38) leads to a contradiction.
2. If ξ = ξ0 e0 + ξ j e j ∈ P j \ Γ with j 6= i : i.e. ξ and y belong to different half-planes, respectively
P j and Pi . Then, from (3.2.36) in Lemma 3.2.5, we have
D(ψ|P j )(ξ) = (C ? + ") f 0 (d(ξ, y))
(ξ0 − y0 )e0 +(ξ j + yi )e j
d(ξ, y)
(C ∗ + ") f 0 (d(ξ, y))
Since
|(ξ0 − y0 )e0 +(ξ j + yi )e j |
d(ξ, y)
and (3.2.32) in Lemma 3.2.4 gives us
|(ξ0 − y0 )e0 + (ξ j + yi )e j |
d(ξ, y)
≤ C ∗.
(3.2.39)
= 1 and f 0 (t) ≥ 1 for all t ∈ [0, 3r), (3.2.39) leads to a contradiction.
3. If ξ = ξ0 e0 ∈ Γ : In this case, ξ and y belong to the same half-plane, but we have to deal with
(3.2.33) in Lemma 3.2.4. For i ∈ {1, . . . , N } be such that y ∈ Pi , from (3.2.36) in Lemma 3.2.5
we have
− yi
≤ 0,
∂ x i ψ|Pi (ξ) = (C ∗ + ") f 0 (d(ξ, y))
d(ξ, y)
and
∂ x 0 ψ(ξ) = (C ∗ + ")g 0 (d(ξ, y))
ξ 0 − y0
.
d(ξ, y)
Then, (3.2.33) in Lemma 3.2.4 gives us
(C ∗ + ") f 0 (d(ξ, y))
Since
|ξ0 − y0 |+ yi
d(ξ, y)
|ξ0 − y0 | + yi
≤ C ∗.
d(ξ, y)
≥ 1 and f 0 (t) ≥ 1 for all t ∈ [0, 3r), (3.2.40) leads to a contradiction.
(3.2.40)
92
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
This conclude the proof of (3.2.37). Remark that if d(x, z) < r then d(x, y) < 2r and f (d(x, y)) =
d(x, y). Then, (3.2.37) yields that
u(x) ≤ u( y) + (C ∗ + ")d(x, y),
∀x, y ∈ S s.t d(x, z) < r, d( y, z) < r.
By symmetry, we get
|u(x) − u( y)| ≤ (C ∗ + ")d(x, y),
∀x, y ∈ S s.t d(x, z) < r, d( y, z) < r,
and by letting " tend to zero, we get
|u(x) − u( y)| ≤ C ∗ d(x, y),
∀x, y ∈ S s.t d(x, z) < r, d( y, z) < r.
(3.2.41)
Now, for two arbitrary points x, y in S ∩ B(Γ , R), we take r = 41 min(R − dist(x, Γ ), R − dist( y, Γ )) and
choose a finite sequence (z j ) j=1,...,M ∈ G belonging to the geodesic between x and y, such that z1 = x,
P M −1
z M = y, d(zi , zi+1 ) < r for all i = 1, . . . , M − 1 and i=1 d(zi , zi+1 ) = d(x, y). From (3.2.41), we get
that
|u(x) − u( y)| ≤ C ∗ d(x, y), ∀x, y ∈ S ∩ B(Γ , R).
An important consequence of this lemma is the following result.
Lemma 3.2.7. Assume [H0], [H1], [H2] and [H3]. Let u be a bounded viscosity subsolution of (3.2.15)
and ϕ : Γ → R a C 1 -function. Then, for any local maximum point x̄ in Γ of u − ϕ on Γ , one has
λu(x̄) + HΓT (x̄, D(ϕ|Γ )(x̄)) ≤ 0.
Proof. Let ϕ : Γ → R be a C 1 -function and x̄ ∈ Γ be a local maximum point of (u − ϕ) |Γ on Γ . Since u is
a subsolution of (3.2.15), according to Lemma 3.2.6, the function u is Lipschitz continuous on B(Γ , r)
for some positive real r, with a Lipschitz constant Lu,r . We introduce the function ϕ defined on S by
ϕ(x 0 e0 + x i ei ) = ϕ(x 0 e0 ) + Lu,r x i . By construction, the function ϕ belongs to R(S ) and u − ϕ admits
a local maximum at the point x̄ on S . Then, since u is a viscosity subsolution of (3.2.15), we see that
λu(x̄) + HΓ (x̄, D(ϕ|P1 )(x̄), . . . , D(ϕ|PN )(x̄)) ≤ 0,
which implies that λu(x̄) + HΓT (x̄, D(ϕ|Γ )(x̄)) ≤ 0.
Remark 3.2.8. The conclusion of Lemma 3.2.7 holds if we replace [H3] with the assumption that the
subsolution u is Lipschitz continuous.
The following lemma can be found in [BBC13, BBC14] in a different context:
Lemma 3.2.8. Assume [H0], [H1], [H2] and [H3]. Let w : S → R be a viscosity supersolution of (3.2.15)
in S . Then if x ∈ Pi \Γ , we have for all t > 0,
‚Z
w(x) ≥ inf
αi (·),θi
t∧θi
`i ( y xi (s), αi (s))e−λs ds
+
w( y xi (t
∧ θi ))e
−λ(t∧θi )
Œ
,
(3.2.42)
0
where αi ∈ L ∞ (0, ∞; Ai ), y xi is the solution of y xi (t) = x +
€R t
0
Š
f i ( y xi (s), αi (s))ds and θi is such that
y xi (θi ) ∈ Γ and θi lies in [τi , τ̄i ], where τi is the exit time of y xi from Pi \Γ and τ̄i is the exit time of y xi
from Pi .
Proof. See [BBC13].
e 3]. See [BBC13].
Remark 3.2.9. The conclusions of Lemma 3.2.8 hold if we replace [H3] with [H
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
93
Remark 3.2.10. Note that the comparison result of Barles-Perthame [BP90] imply the following suboptimality principle for subsolutions that will not be needed in the sequel: let w be a continuous viscosity
subsolution of (3.2.15) in S . If x ∈ Pi \Γ , we have for all t > 0,
‚Z t∧θ
Œ
i
w(x) ≤ inf sup
αi (·) θ
i
`i ( y xi (s), αi (s))e−λs ds + w( y xi (t ∧ θi ))e−λ(t∧θi ) ,
(3.2.43)
0
where αi ∈ L ∞ (0, ∞; Ai ), y xi is the solution of y xi (t) = x +
€R t
0
Š
f i ( y xi (s), αi (s))ds and θi is such that
y xi (θi ) ∈ Γ and θi lies in [τi , τ̄i ], where τi is the exit time of y xi from Pi \Γ and τ̄i is the exit time of y xi
from Pi .
The following theorem is reminiscent of Theorem 3.3 in [BBC13]:
Theorem 3.2.4. Assume [H0], [H1], [H2] and [H3]. Let r > 0 be given by Lemma 3.2.6. Let w : S → R
be a viscosity supersolution of (3.2.15), bounded from below by −c|x| − C for two positive numbers c and
C. Let φ ∈ R(S ) and x̄ ∈ Γ be such that w − φ has a local minimum point at x̄. Then, either [A] or [B]
below is true:
[A] There exists a sequence (ηk )k∈N of positive real numbers such that limk→+∞ ηk = η > 0, an index
i ∈ {1, . . . , N } and a sequence x k ∈ Pi \ Γ such that limk→+∞ x k = x̄ and limk→+∞ w(x k ) =
w(x̄) satisfying the following: for each k, there exists a control law αki such that the corresponding
trajectory y x k (s) ∈ Pi ∩ B(Γ , r) for all s ∈ [0, ηk ] and
w(x k ) ≥
Z
ηk
`i ( y x k (s), αki (s))e−λs ds + w( y x k (ηk ))e−ληk .
(3.2.44)
0
[B]
λw(x̄) + HΓT (x̄, D(φ|Γ )(x̄)) ≥ 0.
(3.2.45)
Proof. For any i in {1, . . . , N } we consider the function ϕi : R → R defined as follows
ϕi (d) = H i (x̄, D φ|Pi (x̄) + dei )
For i in {1, . . . , N }, let qi be a real number such that ϕi (qi ) = mind∈R {ϕi (d)}. We can already remark
that according to Lemma 3.2.1 we have,
HΓ (x̄, D(φ|P1 )(x̄) + q1 e1 , . . . , D(φ|PN )(x̄) + qN eN ) = HΓT (x̄, D(φ|Γ )(x̄)).
Consider the function
ψ" (x) = w(x) − φ(x) − qi x i +
x i2
"2
(3.2.46)
if x ∈ Pi ,
defined on S , where x i = x.ei . Changing φ(x) into φ(x) − |x − x̄|2 if necessary, we can assume that
x̄ is a strict local minimum point of w − φ. Then, standard arguments show that for " small enough,
the function ψ" reaches its minimum close to x̄, and that any sequence of such minimum points x "
converges to x̄ and satisfies w(x " ) converges to w(x̄).
Up to the extraction of a subsequence, we can make out two cases.
1. If x " ∈ Γ : Then, since w is a viscosity supersolution of (3.2.15), we have
λw(x " ) + HΓ x " , D φ|P1 (x " ) + q1 e1 , . . . , D φ|PN + qN eN (x " ) ≥ 0,
and letting " tend to 0, we obtain
λw(x̄) + HΓ x̄, D φ|P1 (x̄) + q1 e1 , . . . , D φ|PN (x̄) + qN eN ≥ 0.
Then, using (3.2.46) we deduce from (3.2.47) that
λw(x̄) + HΓT (x̄, D (φ|Γ ) (x̄)) ≥ 0,
hence [B].
(3.2.47)
94
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
2. If x " ∈ Pi \ Γ for some i ∈ {1, . . . , N } : We skip the writing of this case which is treated as in the
corresponding case in the proof of [AOT15b, Theorem 4.1], using the superoptimality (3.2.42)
of Lemma 3.2.8.
e 3]. Indeed, the proof
Remark 3.2.11. The conclusions of Theorem 3.2.4 hold if we replace [H3] with [H
e 3].
is based on Lemma 3.2.1 and Lemma 3.2.8 which stay true if we replace [H3] with [H
Lemma 3.2.9. Assume [H0],[H1],[H2] and [H3]. Let r > 0 be given by Lemma 3.2.6: any bounded
subsolution of (3.2.15) is Lipschitz continuous in B(O, r) ∩ G . Consider
€R t i ∈ {1, · · · , N }, x Š∈ (Ji \ {O}) ∩
∞
B(O, r), αi ∈ L (0, ∞; Ai ). Let η > 0 be such that y x (t) = x + 0 f i ( y x (s), αi (s))ds ei belongs to
Ji ∩ B(O, r) for any t ∈ [0, η]. For any bounded viscosity subsolution u of (3.2.15) which is a C 1 function
on Γ we have,
Zη
`i ( y x (t), αi (t))e−λt d t + u( y x (η))e−λη .
u(x) ≤
(3.2.48)
0
Proof. Since u is Lipschitz continuous in B(Γ , r)∩Pi , the function t 7−→ u( y x (t)) is Lipschitz continuous
in [0, η]. Let us define the set KΓ = {t ∈ [0, η] : y x (t) ∈ Γ } and KΓc = [0, η] \ KΓ . It is clear that
KΓ is closed and that KΓc an open subset of [0, η]. Let us introduce the function w defined in Pi by
w(x 0 e0 + x i ei ) = u(x 0 e0 ). This function is C 1 in Pi and coincides with the function u on Γ . Then, from
Stampacchia’s theorem,
Zη
Zη
d
d
−λt
1KΓ (t)
1KΓ (t)
u( y x (t))e
dt =
w( y x (t))e−λt d t.
(3.2.49)
dt
dt
0
0
As the function w is C 1 we have
Zη
Zη
d
−λt
1KΓ (t) (Dw( y x (t)). f i ( y x (t), αi (t)) − λw( y x (t))) e−λt d t.
w( y x (t)e
dt =
1KΓ (t)
d
t
0
0
However, since the function w is independent of the variable x i we have
1KΓ (t)Dw( y x (t)) = 1KΓ (t)D (w|Γ ) ( y x (t)).
Then, since u and w coincide on Γ we get that
Zη
Zη
d
−λt
1KΓ (t)
1KΓ (t) (D (u|Γ ) ( y x (t)). f i ( y x (t), αi (t)) − λu( y x (t))) e−λt d t.
w( y x (t))e
dt =
d
t
0
0
(3.2.50)
Since u is a subsolution of (3.2.15) which is C 1 on Γ , we deduce from Lemma 3.2.7 that
λu( y x (t)) + HΓT ( y x (t), D (u|Γ ) ( y x (t))) ≤ 0.
(3.2.51)
However, from Stampacchia’s theorem we have f i ( y x (t), αi (t)).ei = y x0 (t).ei = 0 almost everywhere on
KΓ . Thus, thanks to the definition of HΓT , given by (3.2.21), and (3.2.51) we get
D (u|Γ ) ( y x (t)). f i ( y x (t), αi (t)) − λu( y x (t)) ≥ −`i ( y x (t), αi (t))
for a.a. t ∈ KΓ .
We finally deduce from (3.2.49), (3.2.50) and (3.2.52) that
Zη
Zη
d
u( y x (t))e−λt d t ≥ −
1KΓ (t)
1KΓ (t)`i ( y x (t), αi (t))e−λt d t.
d
t
0
0
(3.2.52)
(3.2.53)
One the other hand, since KΓc is an open subset of [0, η], there exists a countable family of disjoint
intervals (ω j ) j∈J , ω j ⊂ [0, η] such that KΓc = ∪ j∈J w j . Let a j < b j be the lower and upper endpoints of
ω̄ j . From a classical suboptimality principle, see [BCD97, Theorem III.2.33], we see that for any j ∈ J,
u( y x (b j ))e
−λb j
− u( y x (a j ))e
−λa j
Z
bj
`i ( y x (t), αi (t))e−λt d t.
≥−
aj
3.2. FIRST CASE : FULL CONTROLLABILITY NEAR THE INTERFACE
Noting that
u( y x (b j ))e
−λb j
− u( y x (a j ))e
−λa j
=
Z
η
0
95
d
u( y x (t))e−λt 1(a j ,b j ) (t)d t,
dt
and summing over j ∈ J, we obtain that
Zη
Zη
d
−λt
c
1KΓ (t)
u( y x (t))e ) d t ≥ −
`i ( y x (t), αi (t))1KΓc (t)d t.
dt
0
0
(3.2.54)
We get (3.2.48) by summing (3.2.53) and (3.2.54).
Remark 3.2.12. The conclusion of Lemma 3.2.9 holds if we replace [H3] with the assumption that the
subsolution u is a Lipschitz continuous function in a neighborhood Γ which is C 1 on Γ .
3.2.4
Comparison principle and uniqueness
Lemma 3.2.10. Assume [H0], [H1], [H2] and [H3]. Let u : S → R be a bounded continuous viscosity
subsolution of (3.2.15). Let (ρ" )" be a sequence of mollifiers defined on R as follows
ρ" (x) =
where
ρ∈C
∞
Z
+
(R, R ),
1 x
ρ( ),
" "
ρ(x)d x = 1
and supp (ρ) = [−1, 1].
R
We consider the function u" defined on S by
Z
u((x 0 − τ)e0 + x 0 )ρ" (τ)dτ,
u" (x) = u ∗ ρ" (x) =
if x = x 0 e0 + x 0 ,
R
where the decomposition of x ∈ S , x = x 0 e0 + x 0 is explained in (3.2.2).
Then, u" converges uniformly to u in L ∞ (S , R) and there exists a function m : (0, +∞) → (0, +∞), such
that lim"→0 m(") = 0 and the function u" − m(") is a viscosity subsolution of (3.2.15) on a neighborhood
of Γ .
Proof. The uniform convergence of u" to u in L ∞ (S , R) is classical because u is bounded and continuous
in S . The existence of a function m : (0, +∞) → (0, +∞) with m(0+ ) = 0 such that u" − m(")
is a subsolution of (3.2.15) on a neighborhood of Γ was proved by P.-L. Lions [Lio82] or Barles &
Jakobsen [BJ02]. A crucial information for obtaining this result is the fact that, according to Lemma
3.2.6, u is Lipschiz continuous on B(Γ , r) ∩ S , for some positive number r.
Theorem 3.2.5. Assume [H0], [H1], [H2] and [H3]. Let u : S → R be a bounded viscosity subsolution
of (3.2.15), and v : S → R be a bounded viscosity supersolution of (3.2.15). Then u ≤ v in S .
Proof. The strategy of proof adopted here is the one of Barles-Briani-Chasseigne [BBC13] in the proof
of their theorem 4.1. It is adapted to the present context.
Let u" be the approximation of u given by Lemma 3.2.10. It is a simple matter to check that there exists
a positive real number M such that the function ψ(x) = −|x|2 − M is a viscosity subsolution of (3.2.15).
For 0 < µ < 1, µ close to 1, the function u",µ = µ(u" − m(")) + (1 − µ)ψ is a viscosity subsolution of
(3.2.15), which tends to −∞ as |x| tends to +∞. Let M",µ be the maximal value of u",µ − v which is
reached at some point x̄ ",µ . We argue by contradiction assuming that M",µ > 0.
1. If x̄ ",µ ∈
/ Γ , then we introduce the function u",µ (x)− v(x)−d 2 (x, x̄ ",µ ), where d(., .) is the geodesic
distance defined by (3.2.3), which has a strict maximum at x̄ ",µ , and we double the variables, i.e.
for 0 < β 1, we consider
(x, y) 7−→ u",µ (x) − v( y) − d 2 (x, x̄ ",µ ) −
d 2 (x, y)
.
β2
Classical arguments then lead to the conclusion that u",µ (x̄ ",µ )− v(x̄ ",µ ) ≤ 0, thus M",µ ≤ 0, which
is a contradiction.
96
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
2. If x̄ ",µ ∈ Γ , according to Lemma 3.2.6, u",µ is Lipschitz continuous in a neighborhood of Γ . Then,
with a similar argument as in the proof of Lemma 3.2.7, we can construct a test function ϕ ∈
R(S ) such that ϕ|Γ = u",µ |Γ and ϕ remains below u",µ on a neighborhood of Γ (take for exemple
ϕ(x) = u",µ (x 0 , 0) − C x i if x = x 0 e0 + x i ei ∈ Pi , with C great enough). It is easy to check that
v −ϕ has a local minimum at x̄ ",µ , from the assumption that M",µ > 0. Then, we can use Theorem
3.2.4 and we have two possible cases:
[B] λv(x̄ ",µ ) + HΓT (x̄ ",µ , D(u",µ |Γ )(x̄ ",µ )) ≥ 0.
Moreover, u",µ is a subsolution of (3.2.15), C 1 on Γ . Then, according to Lemma 3.2.7, we
have the inequality λu",µ (x̄ ",µ ) + HΓT (x̄ ",µ , D(u",µ |Γ )(x̄ ",µ )) ≤ 0. Therefore, we obtain that
u",µ (x̄ ",µ ) ≤ v(x̄ ",µ ), thus M",µ ≤ 0, which is a contradiction.
[A] With the notations of Theorem 3.2.4, we have that
Z ηk
`i ( y x k (s), αki (s))e−λs ds + v( y x k (ηk ))e−ληk .
v(x k ) ≥
0
Moreover, from Lemma 3.2.9,
Z ηk
u",µ (x k ) ≤
`i ( y x k (s), αki (s))e−λs ds + u",µ ( y x k (ηk ))e−ληk .
0
Therefore
u",µ (x k ) − v(x k ) ≤ (u",µ ( y x k (ηk )) − v( y x k (ηk )))e−ληk .
Letting k tend to +∞, we find that M",µ ≤ M",µ e−λη , which implies that M",µ ≤ 0, which is
a contradiction.
We conclude by letting " tend to 0 and µ tend to 1.
Theorem 3.2.6. Assume [H0], [H1], [H2] and [H3]. Then, the value function v is the unique viscosity
solution of (3.2.15) in S .
3.3
3.3.1
Second case : normal controllability near interface
The new framework
We keep assumptions [H0], [H1], [H2] and we weaken the controllability assumption [H3] by only
supposing normal controllability,
e 3] There is a real number δ > 0 such that for any i = 1, . . . , N and for all x ∈ Γ ,
[H
[−δ, δ] ⊂ { f i (x, a).ei : a ∈ Ai } .
The following property is the counterpart of property 3.2.2 in the current framework.
e 3]. Then, there exists R > 0 a positive real number such that for all
Property 3.3.1. Assume [H0] and [H
i = 1, . . . , N and x ∈ B(Γ , R) ∩ Pi
δ δ
[− , ] ⊂ { f i (x, a).ei : a ∈ Ai } .
2 2
(3.3.1)
The dynamics f and the cost function ` are defined on
M = (x, a); x ∈ S , a ∈ Ai if x ∈ Pi \Γ , and a ∈ ∪Ni=1 Ai if x ∈ Γ ,
as in (3.2.5) and (3.2.9). As above, we need to introduce the set of the admissible controlled trajectories.
For this purpose, we recall that for x ∈ S ,
§
Fi (x)
if x belongs to the open half-plane Pi \Γ
F̃ (x) =
∪Ni=1 Fi (x)
if x ∈ Γ ,
ẏ (t) ∈ F̃ ( y (t)),
for a.a. t > 0
x
Yx = y x ∈ Lip(R+ ; S ) : x
.
(3.3.2)
y x (0) = x,
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
97
e 3]. Then
Theorem 3.3.1. Assume [H0], [H2] and [H
1. For any x ∈ S , Yx is not empty.
2. For any x ∈ S , for each trajectory y x ∈ Yx , there exists a measurable function Φ : [0, +∞) → M ,
Φ(t) = (ϕ1 (t), ϕ2 (t)) with
( y x (t), ẏ x (t)) = (ϕ1 (t), f (ϕ1 (t), ϕ2 (t))),
for a.a. t.
3. Almost everywhere on {t : y x (t) ∈ Γ }, f ( y x (t), ϕ2 (t)) ∈ Re0 .
Points 2 and 3 are proved exactly as in the Theorem 3.2.2. Point 1 is little bit more difficult to prove
e 3], it may happen that 0 6∈ F̃ (x) when x ∈ Γ . Here, point 1 is essentially a consequence
because with [H
of the following lemma.
e 3]. There exists T > 0 such that ∀i ∈ {1, . . . , N } and ∀x ∈ Pi ,
Lemma 3.3.1. Assume [H0], [H2] and [H
there exists y x ∈ Lip([0, T]; Pi ) a solution of the differential inclusion
§
ẏ x (t) ∈ F̃ ( y x (t))
y x (0) = x.
Proof. Take x ∈ Pi .
e 3], there exists ai ∈ Ai such that f i (x, ai ).ei = δ. Then, from the
• If x ∈ Γ : according to [H
Lipschitz property in [H0], we have
f i (z, ai ).ei ≥
δ
,
2
∀z ∈ B(x,
δ
) ∩ Pi .
2L f
(3.3.3)
As a consequence, if we set T = 2L δf M f , there exists a unique function y x : [0, T ) → Pi such that
Rt
y x (t) = x + 0 f i ( y x (s), ai )ds and y x (t) ∈ B(x, 2Lδ f ) ∩ Pi ∀t ∈ (0, T ].
• If x ∈ Pi \ Γ : then, we chose arbitrarily ai ∈ Ai and we consider ȳ : [0, T̄ ) → Pi the maximal
Rt
solution of the integral equation y(t) = x + 0 f i ( y(s), ai )ds. If T̄ > T = 2L δf M f then we take
y x = ȳ. Otherwise ȳ( T̄ ) is well defined and belongs to Γ , then we are reduced to the case where
x ∈ Γ.
Now, we are able to prove Point 1 of Theorem 3.3.1.
Proof. Let x be in Pi , for some i ∈ {1, . . . , N }. According to Lemma 3.3.1 we are able to build a sequence
of positive reals (Tn )n∈N and maps yn : [0, Tn ] → Pi such that, for any n ∈ N
(i) yn (0) = x and ẏn (t) ∈ Fi ( yn (t)), for a.a. t ∈ (0, Tn ]
(ii) yn (t) ∈ Pi , ∀t ∈ (0, Tn ]
(iii) Tn+1 − Tn ≥ T , with T given by Lemma 3.3.1
(iv) ∀0 ≤ q ≤ p, yq |[0,Tp ] ≡ y p
So, the map y x : R+ → Pi defined by y x (t) = yn (t) if t ∈ [0, Tn ] belongs to Yx and particularly Point 1
of Theorem 3.3.1 is true.
x:
So, as above, we can take the set of admissible controlled trajectories starting from the initial datum


y x ∈ Lip(R+ ; S ),


Z t
∞
T x = ( y x , α) ∈ LLoc
(R+ ; M ) : .
+
f ( y x (s), α(s))ds in R 

y x (t) = x +
0
98
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Then, the cost functional J and the value function v are defined by
J(x; ( y x , α)) =
Z
∞
`( y x (t), α(t))e−λt d t,
0
where λ > 0, and
v(x) =
inf
( y x ,α)∈T x
J(x; ( y x , α)).
Unlike in § 3.2, we cannot use classical arguments to prove that v is continuous, because we do not
e 3] we are no longer sure that for each x, z
suppose [H3] any longer. The main problem is that with [H
close to Γ , there exists an admissible trajectory y x,z ∈ T x from x to z. We will prove later that v is
continuous, but for the moment v is a priori a discontinuous function. In order to deal with this a priori
discontinuity, we use the following notions :
Definition 3.3.1. Let u : S → R be a bounded function.
• The lower semi-continuous envelope of u is defined by
u? (x) = lim inf u(z).
z→x
• The upper semi-continuous envelope of u is defined by
u? (x) = lim sup u(z).
z→x
3.3.2
Hamilton-Jacobi equation
Definition 3.3.2. A bounded function u : S → R is a discontinuous viscosity solution of (3.2.15) in S if
u? is a subsolution and u? is a supersolution of (3.2.15) in S .
The next lemmas will be used to prove the existence result.
e 3].There exists some constants r > 0 and C > 0 such that
Lemma 3.3.2. Assume [H0], [H1], [H2] and [H
for all x = x 0 e0 + x i ei ∈ B(Γ , r) ∩ (Pi \ Γ ), there exists an admissible controlled trajectory ( y x , α x ) ∈ T x
such that τ x ≤ C x i , where τ x is the exit time of y x from Pi \ Γ .
e 3]. For all x ∈ Γ , i ∈ {1, . . . , N } and a ∈ Ai such that
Lemma 3.3.3. Assume [H0], [H1], [H2] and [H
f i (x, a).ei ≥ 0, there exists a sequence a" ∈ Ai such that
f i (x, a" ).ei
≥
δ" > 0,
(3.3.4)
| f i (x, a" ) − f i (x, a)|
≤
2M f ",
(3.3.5)
|`i (x, a" ) − `i (x, a)|
≤
2M` ".
(3.3.6)
e 3] there exists aδ ∈ Ai such that f i (x, aδ ).ei = δ. From [H2],
Proof. From [H
"( f i (x, aδ ), `i (x, aδ )) + (1 − ")( f i (x, a), `i (x, a)) ∈ FLi (x)
for any " ∈ [0, 1]. So, there exists a" ∈ Ai such that
"( f i (x, aδ ), `i (x, aδ )) + (1 − ")( f i (x, a), `i (x, a)) = ( f i (x, a" ), `i (x, a" ))
which has all the desired properties.
Corollary 3.3.1. For any i ∈ {1, . . . , N }, x ∈ Γ and pi ∈ Re0 × Rei ,
a∈Ai
max
(−pi f i (x, a) − `i (x, a)) =
sup
(−pi f i (x, a) − `i (x, a)).
s.t. f i (x,a).ei ≥0
a∈Ai s.t. f i (x,a).ei >0
(3.3.7)
e 3]. The value function v defined in (3.2.10) is a bounded
Theorem 3.3.2. Assume [H0], [H1], [H2] and [H
discontinuous viscosity solution of (3.2.15) in S .
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
99
Proof. The proof being a bit long, we postpone it to the appendix 3.5.5.
We give now some results on the behavior of the Hamiltonians in the new framework.
e 3]. Then, for i = 1, . . . , N , we have
Lemma 3.3.4. Assume [H0], [H1], [H2] and [H
|H i (x, p) − H i ( y, p)| ≤ L f |x − y||p| + ω` (|x − y|),
for any x, y ∈ Pi and p ∈ Re0 × Rei ,
(3.3.8)
and
|H i (x, p) − H i (x, q)| ≤ M f |p − q|,
for any x ∈ Pi and p, q ∈ Re0 × Rei ,
(3.3.9)
for any x ∈ B(Γ , R) ∩ Pi and p0 , pi ∈ R,
(3.3.10)
where L f , M f and ω` are defined in [H0] and [H1].
At last, if C M = max{M f , Ml },
H i (x, p0 e0 + pi ei ) ≥
δ
|pi | − C M (1 + |p0 |),
2
where R > 0 is a positive number as in (3.3.1).
Proof. The proof of the Lemma is given in [BBC13] Lemma 7.1 We supply it for completeness. Assumptions [H0] and [H1] yield (3.3.8) and (3.3.9). For (3.3.10) we also have to use the partial controllability
e 3]. Indeed, according to Property 3.3.1 there exist some controls a1 , a2 ∈ Ai such that
assumption [H
− f i (x, a1 ).ei =
δ
> 0,
2
δ
− f i (x, a2 ).ei = − .
2
Now we compute H i (x, p0 e0 + pi ei ) assuming that pi > 0 (the other case is treated similarly).
H i (x, p0 e0 + pi ei )
≥
≥
≥
− f i (x, a1 ).(p0 e0 + pi ei ) − `i (x, a1 )
δ
2 |pi | − f i (x, a1 ).p0 e0 − `i (x, a1 )
δ
2 |pi | − C M |p0 | − C M ,
the last line coming from the boundedness of f i and l i . This concludes the proof.
e 3]. Then, for any x, x 0 ∈ Γ , i ∈ {1, . . . , N } and a ∈ Ai
Lemma 3.3.5. Assume [H0], [H1], [H2] and [H
0
such that f i (x, a).ei ≥ 0, there exists a ∈ Ai such that f i (x 0 , a0 ).ei ≥ 0 and
| f i (x, a) − f i (x 0 , a0 )| ≤ M |x − x 0 |,
0
0
0
|`i (x, a) − `i (x , a )| ≤ ω(|x − x |),
(3.3.11)
(3.3.12)
e 3].
where, if M f , L f , M` , δ and ω` are given by assumptions [H0], [H1] and [H
M = L f 1 + 2M f δ−1
and ω(t) = ω` (t) + 2M` L f δ−1 t for t ≥ 0.
Proof. The proof follows the lines of that of [BBC13], Lemma 6.6 : let a ∈ Ai be such that f i (x, a).ei ≥ 0.
Fix x 0 ∈ Γ , we have two possibilities. If f i (x 0 , a).ei ≥ 0 the conclusion follows easily because according
to [H0] and [H1] we have respectively (3.3.11) and (3.3.12) with a0 = a. Otherwise f i (x 0 , a).ei < 0. By
e 3] there exists a control a1 ∈ Ai such that f i (x 0 , a1 ).ei = δ.
the partial controllability assumption in [H
We then set
δ
.
µ :=
δ − f i (x 0 , a).ei
Since µ ∈ (0, 1), by the convexity assumption in [H2], there exists a control a0 ∈ Ai such that
µ f i (x 0 , a), `i (x 0 , a) + (1 − µ) f i (x 0 , a1 ), `i (x 0 , a1 ) = f i (x 0 , a0 ), `i (x 0 , a0 ) .
By construction f i (x 0 , a0 ).ei = 0 and then, in particular, f i (x 0 , a0 ).ei ≥ 0. Moreover, since
(1 − µ) =
we have
L f |x − x 0 |
f i (x, a).ei − f i (x 0 , a).ei
− f i (x 0 , a).ei
≤
≤
,
δ − f i (x 0 , a).ei
δ − f i (x 0 , a).ei
δ
| f i (x, a) − f i (x 0 , a0 )|
≤
≤
≤
| f i (x, a) − f i (x 0 , a)| + | f i (x 0 , a) − f i (x 0 , a0 )|
L f |x − x 0 | + (1 − µ̄)| f i (x 0 , a) − f i (x 0 , a1 )|
L f 1 + 2M f δ−1 |x − x 0 |.
This proves (3.3.11). The same calculation with `i gives us (3.3.12).
100
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Remark 3.3.1. In Lemma 3.3.5, if x ∈ Γ and a ∈ Ai are such that f i (x, a).ei = 0, then we have the
stronger result that ∀x 0 ∈ Γ , ∃a0 ∈ Ai such that f i (x 0 , a0 ).ei = 0 and the inequalities (3.3.11) and (3.3.12)
are true. Indeed, if f i (x 0 , a).ei ≤ 0 the proof of Lemma 3.3.5 directly provides a suitable a0 ∈ Ai . We
e 3]
only need to specify the strategy in the case when f i (x 0 , a).ei > 0. We consider a2 ∈ Ai , given by [H
δ
0
such that f i (x 0 , a2 ).ei = −δ. Then, we set µ̃ = f i (x 0 ,a).e
∈
(0,
1)
and
we
choose
a
∈
A
,
given
by
[H2],
i
i +δ
such that µ̃ ( f i (x 0 , a), `i (x 0 , a)) + (1 − µ̃) ( f i (x 0 , a2 ), `i (x 0 , a2 )) = ( f i (x 0 , a0 ), `i (x 0 , a0 )). With this choice
f (x 0 ,a).ei
( f (x 0 ,a)− f i (x,a)).ei
for a0 ∈ Ai , the same calculations as in the proof above, using 1 − µ̃ = f i (xi 0 ,a).ei +δ
= i f i (x 0 ,a).e
, allow
i +δ
us to conclude.
e 3]. The Hamiltonian H + defined in (3.2.20) has the
Lemma 3.3.6. Assume [H0], [H1], [H2] and [H
i
following properties
H i+ (x, p0 e0 + pi ei ) ≥ H i+ (x, p0 e0 + qi ei ) ∀x ∈ Γ and ∀p0 , pi , qi ∈ R s.t pi ≤ qi ,
H i+ (x, p0 e0 + pi ei ) ≥ −δpi − C M (1 + |p0 |),
∀x ∈ Γ and p0 , pi ∈ R,
|H i+ (x, p) − H i+ ( y, p)| ≤ M |x − y||p| + ω(|x − y|) ∀x, y ∈ Γ and p ∈ Re0 × Rei ,
(3.3.13)
(3.3.14)
(3.3.15)
where M and ω are defined in Lemma 3.3.5 and C M = max{M f , M` }.
Proof. We skip the easy proof of (3.3.13). The proof of (3.3.14) is the same as that of (3.3.10) with
the difference that here the dynamics leading out of Pi are not allowed. Let us prove (3.3.15). Let
x, y be in Γ and p be in Re0 × Rei . First, there exists a ∈ Ai such that f i (x, a).ei ≥ 0 and H i+ (x, p) =
− f i (x, a).p − `i (x, a). Then, we consider a0 ∈ Ai given by Lemma 3.3.5 such that f i ( y, a0 ).ei ≥ 0 and
the inequality (3.3.11) and (3.3.12) are satisfied. Therefore,
H i+ (x, p) − H i+ ( y, p)
≤
=
≤
(− f i (x, a).p − `i (x, a)) − (− f i ( y, a0 ).p − `i ( y, a0 ))
( f i ( y, a0 ) − f i (x, a)) .p + (`i ( y, a0 ) − `i (x, a))
M |x − y||p| + ω(|x − y|).
We conclude the proof by exchanging the roles of x and y.
Remark 3.3.2. In view of the calculations above, if the maximum which defines H i+ (x, p0 e0 + pi ei ) is
reached for some a ∈ Ai such that f i (x, a).ei = 0, then from the Remark 3.3.1 we have the stronger
inequality,
H i+ (x, p0 e0 + pi ei ) − H i+ ( y, p0 e0 + pi ei ) ≤ M |x − y||p0 | + ω(|x − y|),
(3.3.16)
but without the modulus on the left side of the inequality.
3.3.3
Comparison principle and Uniqueness
A key argument in the proof of the Comparison Principle in Theorem 3.2.5 is the fact that the subsolutions of (3.2.15) are Lipschitz continuous in a neighborhood of Γ . We have not this property in the
current framework the above method cannot be applied directly. We will first prove a local comparison
principle by reducing ourselves to the case when a subsolution is Lipschitz continuous, and then we
will deduce a global comparison principle. For this purpose, we start by stating some useful lemmas.
The following transformation will allow us to focus on the case when the subsolutions are locally
Lipschitz continuous in a neighborhood of Γ .
Definition 3.3.3. Let u : S → R be a bounded, usc function and α be a positive number. We define the
sup-convolution of u with respect to the x 0 -variable by
(
2p )
|z0 − x 0 |2
uα (x) := max u(z0 e0 + x i ei ) −
+α
,
if x = x 0 e0 + x i ei ∈ Pi ,
(3.3.17)
z0 ∈R
α2
where α, p > 0 are positive numbers.
We recall well known results on sup-convolution.
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
101
Lemma 3.3.7. Let u : S → R be a bounded function and α, p be positive numbers. Then, for any x ∈ S
the supremum which defines uα (x) is achieved at a point z0 ∈ R such that
(
p
p
|z0 − x 0 |2
+ α) 2 ≤ 2 k u k∞ +α 2 .
α2
(3.3.18)
Lemma 3.3.8. Let u : S → R be a bounded, usc function. Then, for all α, p > 0, the sup-convolution uα
is locally Lipschitz continuous with respect to x 0 ; i.e. for any compact subset K ⊂ R3 , there exists CK > 0
such that for all x = x 0 e0 + x i ei , y = y0 e0 + x i ei ∈ K ∩ S , |uα (x) − uα ( y)| ≤ CK |x 0 − y0 |.
e 3]. Let R > 0 be a positive number as in (3.3.1). Let
Lemma 3.3.9. Assume [H0], [H1], [H2] and [H
u : S → R be a bounded, usc subsolution of (3.2.15) in S and α, p be some positive numbers. Then, for all
M > 0, uα is Lipschiz continuous in B M (Γ , R)∩S , where B M (Γ , R) := {x = x 0 e0 + x 0 ∈ B(Γ , R) : |x 0 | ≤ M }.
Proof. Fix M > 0. To get that uα is Lipschitz continuous in B M (Γ , R)∩S , it is enough to show that there
exists a positive number C ? (M , α, p), which can depend to M , α and p, such that uα is a subsolution of
|Du(x)| ≤ C ? (M , α, p) in B M (Γ , R) ∩ S ,
(3.3.19)
i.e. for any x ∈ B M (Γ , R) and ϕ ∈ R(S ), such that uα − ϕ has a local maximum point at x,
max
i=1,...,N
|D(ϕ|Pi )(x)| ≤ C ? (M , α, p)
|∂ x 0 ϕ(x)| + ∂ x i (ϕ|Pi )(x) − ≤ C ? (M , α, p)
if x ∈ (B M (Γ , R) ∩ Pi ) \ Γ ,
(3.3.20)
if x ∈ B M (Γ , R) ∩ Γ ,
(3.3.21)
where for x ∈ R, [x]− = max{0, −x}.
Indeed, if (3.3.19) is proved, then the method used in the proof of Lemma 3.2.6 allows us to get that
uα is Lipschitz continuous with Lipschitz constant C ? (M , α, p).
Let us prove the existence of the constant C ? (M , α, p).
According to Lemma 3.3.8, there exists a constant C(M , α, p) such that uα is C(M , α, p)-Lipschitz continuous with respect to the x 0 -variable. A direct consequence of this is that uα is a subsolution of
|∂ x 0 u(x)| ≤ C(M , α, p) in B M (Γ , R) ∩ S .
(3.3.22)
i.e. for any x ∈ B M (Γ , R) ∩ S and ϕ ∈ R(S ), such that uα − ϕ has a local maximum point at x
|∂ x 0 ϕ(x)| ≤ C(M , α, p).
It remains to get information on |∂ x i (uα |Pi )| in a viscosity sense. Let x̄ = x̄ 0 e0 + x̄ i ei ∈ B M (Γ , R) ∩ Pi
and ϕ ∈ R(S ) be such that uα − ϕ has a local maximum at x̄. According to Lemma 3.3.7, there exists
z̄0 ∈ R such that
2p
|z̄0 − x̄ 0 |2
uα (x̄) := u(z̄0 e0 + x̄ i ei ) −
+α
(3.3.23)
α2
and
2p
p
|z̄0 − x̄ 0 |2
+ α ≤ 2 k u k∞ +α 2 .
2
α
|z −x̄ |2
(3.3.24)
p
Then, if we set ϕα : z0 e0 + x i ei 7−→ ( 0 α2 0 + α) 2 + ϕ(x̄ 0 e0 + x i ei ), we can check that ϕα belongs to
R(S ) and that u − ϕα has a local maximum at z̄ = z̄0 e0 + x̄ i ei . Since u is a subsolution of (3.2.15) in
S , we deduce that
λu(z̄) + sup {−Dϕα (z̄, ζ) − ξ} ≤ 0.
(3.3.25)
(ζ,ξ)∈FL(z̄)
If x̄ ∈ (B M (Γ , R) ∩ Pi ) \ Γ : The equation (3.3.25) can be rewritten as follows
p(z̄0 − x̄ 0 )
λu(z̄) + H i (z̄0 e0 + x̄ i ei ,
α2
2p −1
|x̄ 0 − z̄0 |2
+
α
e
+
∂
ϕ|
(x̄ 0 e0 + x̄ i ei )ei ) ≤ 0.
0
x
P
i
i
α2
(3.3.26)
102
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Then, according to (3.3.10) in Lemma 3.3.4
p|z̄0 − x̄ 0 |
δ
λu(z̄) + |∂ x i ϕ|Pi (x̄)| − C M 1 +
2
α2
But from (3.3.24), the term
K(α, p) =
v
2
t
pp
p
2kuk∞ +α 2
−α
α
p|z̄0 −x̄ 0 |
α2
€
|x̄ 0 −z̄0 |2
α2
p
2 k u k∞ +α 2
2p −1 !
|x̄ 0 − z̄0 |2
+α
≤ 0.
α2
(3.3.27)
Š 2p −1
+α
is bounded by a constant
1− 2p
, so that we get
2
|∂ x i ϕ|Pi (x̄)| ≤ (λ k u k∞ +C M (1 + K(M , α, p))) .
δ
If x̄ ∈ B M (Γ , R) ∩ Γ : equation (3.3.25) can be rewritten as follows

€
Š 2p −1
p|z̄ −x̄ | |x̄ 0 −z̄0 |2
+
α
λu(z̄) + HΓ z̄, 0α2 0
e0 + ∂ x 1 ϕ|P1 (x̄)e1 , . . .
α2
‹
€
Š 2p −1
p|z̄ −x̄ | |x̄ 0 −z̄0 |2
(x̄)e
. . . , 0α2 0
ϕ|
+
α
e
+
∂
2
N ≤ 0.
PN
0
xN
α
(3.3.28)
(3.3.29)
Particularly, if we fix i ∈ {1, . . . , N }, (3.3.29) implies
λu(z̄) + H i+
p|z̄0 − x̄ 0 |
z̄,
α2
!
2p −1
|x̄ 0 − z̄0 |2
e0 + ∂ x i ϕ|Pi (x̄)ei ≤ 0,
+α
α2
(3.3.30)
and thanks to (3.3.14) in Lemma 3.3.6 and (3.3.24) we deduce that
1
− ∂ x i ϕ|Pi (x̄) ≤ (λ k u k∞ +C M (1 + K(M , α, p))) ,
δ
(3.3.31)
with the same constant K(M , α, p) as in the case where x̄ ∈ B M (Γ , R) ∩ Pi .
Finally, according to (3.3.22), (3.3.28) and (3.3.31)
q we have that uα is a subsolution of (3.3.19) in
B M (Γ , R) ∩ S with the constant C ? (M , α, p) =
This concludes the proof.
C(M , α, p)2 +
4
δ2
(λ k u k∞ +C M (1 + K(M , α, p)))2 .
e 3] and [H4]. Let y0 be in Γ and R > 0 be as in (3.3.1). We
Lemma 3.3.10. Assume [H0], [H1], [H2], [H
denote by Q the set S ∩ B( y0 , R). Let u : S → R be a bounded, usc subsolution of (3.2.15) in Q. Then,
for all α, p > 0 small enough, if we set
n
o
q
p
Q α := x ∈ Q : dist(x, ∂ Q) > α (2 k u k∞ +α 2 )2/p − α ,
(3.3.32)
the sup-convolution uα defined in (3.3.17), is Lipschitz continuous in Q α and there exists m : (0, +∞) →
(0, +∞) such that limα→0 m(α) = 0 and uα − m(α) is a subsolution of (3.2.15) in Q α .
Proof. First, according to Lemma 3.3.9 it is clear that uα is Lipschitz continuous in Q α . It remains to
find m : R+ → R+ such that uα − m(α) is a subsolution of (3.2.15) in Q α and to check that m has the
desired limit as α tends to 0. For this purpose, we consider ϕ ∈ R(S ) such that uα − ϕ has a local
maximum point at x̄ = x̄ 0 e0 + x̄ i ei . According to Lemma 3.3.7, there exists z̄0 ∈ R such that
2p
|z̄0 − x̄ 0 |2
uα (x̄) = u(z̄0 e0 + x̄ i ei ) −
+α
α2
(3.3.33)
and
Ç
p 2
|x̄ 0 − z̄0 | ≤ α (2 k u k∞ +α 2 ) p − α.
(3.3.34)
p
€
Š
2
|z̄ −x |2
From (3.3.33) the function x 0 7→ u(z̄0 e0 + x̄ i ei ) − 0 α2 0 + α − ϕ(x 0 e0 + x̄ i ei ) has a local maximum
at x̄ 0 and we deduce that
p(z̄0 − x̄ 0 )
∂ x 0 ϕ(x̄) =
α2
2p −1
|z̄0 − x̄ 0 |2
+α
.
α2
(3.3.35)
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
103
On the other hand, since x̄ ∈ Q α , the inequality (3.3.34) implies that z̄ = z̄0 e0 + x̄ i ei ∈ Q and we can
use that u is a subsolution of (3.2.15) in Q :
1. If x̄ ∈ Pi \ Γ : then, z̄ = z̄0 e0 + x̄ i ei also belongs to Pi \ Γ and using the test function ϕ̃ : z =
€
Š 2p
|z −x̄ |2
z0 e0 + z 0 7−→ ϕ(x̄ 0 e0 + z 0 ) + 0 α2 0 + α , where we recall that z 0 denotes z j e j if z ∈ P j , we have the
viscosity inequality
p(z̄0 − x̄ 0 )
z̄,
α2
λu(z̄) + H i
!
2p −1
|z̄0 − x̄ 0 |2
+α
e0 + ∂ x i ϕ|Pi (x̄)ei ≤ 0.
α2
(3.3.36)
Combining the previous results, (3.3.33), (3.3.35) and (3.3.36), we get
2p
|z̄0 − x̄ 0 |2
+ α + H i (z̄, D ϕ|Pi (x̄)) ≤ 0.
λuα (x̄) + λ
2
α
Then, according to (3.3.8) we have
λuα (x̄)+H i (x̄, D ϕ|Pi
2p
|z̄0 − x̄ 0 |2
(x̄)) ≤ −λ
+
α
+
L
|x̄
−z̄
||D
ϕ|
(x̄)|+ω` (|x̄ 0 −z̄0 |). (3.3.37)
f
0
0
P
i
α2
But, from (3.3.10) and (3.3.36)
|∂ x i ϕ|Pi
2
(x̄)| ≤
δ
λ k u k∞ +C M
p|z̄0 − x̄ 0 |
1+
α2
2p −1 !!
|z̄0 − x̄ 0 |2
.
+α
α2
Then, from (3.3.35) and (3.3.37) we get
λuα (x̄) + H i (x̄, D ϕ|Pi (x̄)) ≤
Šp
2 −1
|z̄0 −x̄ 0 |2
+α
α2
2L
+|x̄ 0 − z̄0 | δ f (λ k u k∞
|z̄0 −x̄ 0 |2
α2
€
pL f 1 +
2C M
δ
−λ
+C M ) + ω` (|x̄ 0 − z̄0 |),
and for p small enough, we get
2L
λuα (x̄) + H i (x̄, D ϕ|Pi (x̄)) ≤ |x̄ 0 − z̄0 | δ f (λ k u k∞ +C M ) + ω` (|x̄ 0 − z̄0 |).
Finally, according to (3.3.34) the right hand side of the latter inequality gives us m(α).
2. If x̄ ∈ Γ : x̄ i = 0 and z̄ = z̄0 e0 + x̄ i ei also belongs to Γ . Using ϕ̃ : z = z0 e0 + z 0 7−→ ϕ(x̄ 0 e0 + z 0 ) +
€
Š 2p
|z0 −x̄ 0 |2
+
α
as a test function, we have
α2

€
Š 2p −1
p(z̄ −x̄ ) |z̄0 −x̄ 0 |2
λu(z̄) + HΓ z̄, 0α2 0
+
α
e0 + ∂ x 1 ϕ|P1 (x̄)e1 , . . .
α2
‹
€
Š 2p −1
p(z̄ −x̄ ) |z̄0 −x̄ 0 |2
e
+
∂
. . . , 0α2 0
+
α
ϕ|
(x̄)e
≤ 0.
2
0
x
P
N
α
N
N
(3.3.38)
As in the case when x̄ ∈ Pi \ Γ , we deduce from (3.3.33), (3.3.35) and (3.3.38) that we have
2p
|z̄0 − x̄ 0 |2
λuα (x̄) + λ
+ α + HΓ (z̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) ≤ 0.
2
α
(3.3.39)
Let i ∈ {1, . . . , N } be such that HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) = H i+ (x̄, D ϕ|Pi (x̄)). Then,
Lemma 3.3.6 and (3.3.39) give us
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄))
≤
Š 2p
€
|z̄ −x̄ |2
−λ 0 α2 0 + α + M |x̄ 0 − z̄0 ||D ϕ|Pi (x̄)|
+ω(|x̄ 0 − z̄0 |),
(3.3.40)
104
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
where M and ω(.) are specified in Lemma 3.3.6. On the other hand, from (3.3.14) in Lemma 3.3.6 and
(3.3.38)
2p −1 !!
p|z̄0 − x̄ 0 | |z̄0 − x̄ 0 |2
1
λ k u k∞ +C M 1 +
+α
− ∂ x i ϕ|Pi (x̄) ≤
.
(3.3.41)
δ
α2
α2
This information does
not allow us to conclude directly with (3.3.40). We need to get a control from
above on ∂ x i ϕ|Pi (x̄). For this purpose, we introduce the real number p̃i,x̄,α defined as follow
p̃i,x̄,α := max{pi ∈ R : ϕi,x̄,α (pi ) = min{ϕi,x̄,α (d)}}
d∈R
where
p(z̄0 − x̄ 0 )
ϕi,x̄,α (d) = H i (x̄,
α2
(3.3.42)
2p −1
|z̄0 − x̄ 0 |2
+α
e0 + dei ).
α2
We claim that if we note hi,x̄ := mind∈R {H i (x̄, dei )}, we have the following estimation
2p −1 !
2p −1 !
p|z̄0 − x̄ 0 | |z̄0 − x̄ 0 |2
p|z̄0 − x̄ 0 | |z̄0 − x̄ 0 |2
2
p̃i,x̄,α ≤
hi,x̄ + C M 1 +
+α
+α
+ Mf
.
δ
α2
α2
α2
α2
(3.3.43)
To prove this inequality, it is enough
to
show
that
for
any
real
number
q
larger
than
the
right
member
of (3.3.43), ϕi,x̄,α (q) > mind∈R ϕi,x̄,α (d) .


Š 2p −1 ‹
Š 2p −1 ‹
€
€
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
2
+α
+α
+ M f α2
. Then, from the
Take q > δ hi,x̄ + C M 1 + α2
α2
α2
coercivity property of H i with respect to pi , see (3.3.10), we have

Š 2p −1
Š 2p −1 ‹
€
€
p(z̄ −x̄ ) |z̄0 −x̄ 0 |2
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
δ
+
α
+
α
H i (x̄, 0α2 0
e
+
qe
)
≥
|q|
−
C
1
+
0
i
M
2
α2
α2
α2
€
Š 2p −1
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
> hi,x̄ + M f α2
+
α
.
2
α
(3.3.44)
But, if we consider d ∈ R such that H i (x̄, d ei ) = hi,x̄ , by definition of pi,x̄,α we have
p(z̄0 − x̄ 0 )
H i (x̄,
α2
2p −1
2p −1
|z̄0 − x̄ 0 |2
p(z̄0 − x̄ 0 ) |z̄0 − x̄ 0 |2
+α
e0 + p̃i,x̄,α ei ) ≤ H i (x̄,
+α
e0 + dei ),
α2
α2
α2
(3.3.45)
and from (3.3.9)
p(z̄0 − x̄ 0 )
H i (x̄,
α2
2p −1
2p −1
|z̄0 − x̄ 0 |2
p|z̄0 − x̄ 0 | |z̄0 − x̄ 0 |2
+α
e0 + dei ) ≤ hi,x̄ + M f
+α
. (3.3.46)
α2
α2
α2
Finally, (3.3.45) and (3.3.46) lead to
p(z̄0 − x̄ 0 )
H i (x̄,
α2
2p −1
2p −1
|z̄0 − x̄ 0 |2
p|z̄0 − x̄ 0 | |z̄0 − x̄ 0 |2
+α
e0 +pi,x̄,α ei ) ≤ hi,x̄ +M f
+α
, (3.3.47)
α2
α2
α2
and from (3.3.44) we infer that
p(z̄0 − x̄ 0 )
H i (x̄,
α2
2p −1
2p −1
|z̄0 − x̄ 0 |2
p(z̄0 − x̄ 0 ) |z̄0 − x̄ 0 |2
+α
e0 + qei ) > H i (x̄,
+α
e0 + p̃i,x̄,α ei ),
α2
α2
α2
which proves (3.3.43).
We consider three
cases :
(a) If ∂ x i ϕ|Pi (x̄) ≤ 0: then, (3.3.41) yields
|∂ x i ϕ|Pi
1
(x̄)| ≤
δ
λ k u k∞ +C M
p|z̄0 − x̄ 0 |
1+
α2
2p −1 !!
|z̄0 − x̄ 0 |2
+α
,
α2
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
105
and therefore (3.3.35) and (3.3.40) imply
€
Š 2p
|z̄ −x̄ |2
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) ≤ −λ 0 α2 0 + α + ω(|x̄ 0 − z̄0 |)


•
€
Š 2p −1
€
Š 2p −1 ‹‹˜
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
p|z̄0 −x̄ 0 | |z̄0 −x̄ 0 |2
1
1
+
λ
k
u
k
+C
+|x̄ 0 − z̄0 |M
+
α
+
+
α
,
∞
M
δ
α2
α2
α2
α2
and then
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄))
≤
Finally, if p is small enough, (3.3.48) implies
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄))
Šp
2
|z̄0 −x̄ 0 |2
+α
α2
+M
δ |z̄0 − x̄ 0 | (λ
€
C
−λ + pM 1 + δM
k u k∞ +C M ) + ω(|x̄ 0 − z̄0 |).
(3.3.48)
M
δ |z̄0
− x̄ 0 | (λ k u k∞ +C M ) + ω(|x̄ 0 − z̄0 |),
(3.3.49)
and from (3.3.34) the
right
hand
side
of
the
latter
inequality
gives
us
m(α).
(b) If 0 < ∂ x i ϕ|Pi (x̄) ≤ p̃i,x̄,α (case which never occurs if pi,x̄,α ≤ 0): in this case, (3.3.35) and
(3.3.40) give us
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄),
≤
Š 2p
€
|z̄ −x̄ |2
. . . , D ϕ|PN (x̄)) ≤ −λ 0 α2 0 + α
€
Š 2p −1
p|z̄ −x̄ |2 |z̄0 −x̄ 0 |2
+
α
+M 0α2 0
+ M |x̄ 0 − z̄0 |p̃i,x̄,α + ω(|x̄ 0 − z̄0 |),
2
α
and according to (3.3.43) we get
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) ≤ 2M
δ |x̄ 0 −p z̄0 |(hi,x̄ + C M ) + ω(|x̄ 0 − z̄0 |)
€
Š2
2
|z̄ −x̄ |
−λ + pM 1 + δ2 C M + M f
.
+ 0 α2 0 + α
Finally, for p small enough, (3.3.50) gives us
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄))
(3.3.50)
2M
δ |x̄ 0
− z̄0 |(hi,x̄ + C M ) + ω(|x̄ 0 − z̄0 |),
(3.3.51)
and since hi,x̄ is uniformly bounded with respect to x̄ ∈ Q, from (3.3.34) the right hand side of the
latter inequality gives us m(α).
(c) If 0 ≤ max{0, p̃i,x̄,α } < ∂ x i ϕ|Pi (x̄): in this case, we can not conclude with the inequality (3.3.40).
We need to find a more precise estimate.
We recall that i ∈ {1, . . . , N } is such that HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) = H i+ (x̄, D ϕ|Pi (x̄)).
Since p̃i,x̄,α < ∂ x i ϕ|Pi (x̄), according to point 4 in Lemma 3.2.1, we know that the maximum which
defines H i+ (x̄, D(ϕ|Pi ))(x̄) is reached for one a ∈ Ai such that f i (x̄, a).ei = 0. Then, we can apply
(3.3.16) in Remark 3.3.2 and (3.3.39) implies
≤
€
Š 2p
|z̄ −x̄ |2
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) ≤ −λ 0 α2 0 + α + ω(|x̄ 0 − z̄0 |)
€
Š 2p −1
p|z̄ −x̄ |2 |z̄0 −x̄ 0 |2
+
α
.
+M 0α2 0
2
α
Then, for p small enough, we deduce that
λuα (x̄) + HΓ (x̄, D ϕ|P1 (x̄), . . . , D ϕ|PN (x̄)) ≤ ω(|x̄ 0 − z̄0 |),
(3.3.52)
and according to (3.3.34) the right hand side of the latter inequality gives us m(α).
Finally, let us state the counterpart of Lemma 3.2.10 in the present context.
Lemma 3.3.11. Assume [H0], [H1] and [H2]. For y0 ∈ Γ let R > 0 be as in (3.3.1). We set Q = B( y0 , R)∩
S . Let u : S → R be a bounded, Lipschitz continuous viscosity subsolution of (3.2.15) in Q. Let (ρ" )" be a
sequence of mollifiers defined on R. We consider the function u" defined on Q " := {x ∈ Q : dist(x, ∂ Q) > "}
by
Z
u" (x 0 e0 + x 0 ) = u ∗ ρ" (x 0 e0 + x 0 ) =
u((x 0 − τ)e0 + x 0 )ρ" (τ)dτ.
R
106
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
We recall that the decomposition of x ∈ S , x = x 0 e0 + x 0 , is explained in (3.2.2).
Then, k u" − u k L ∞ (Q " ) tends to 0 as " tends to 0 and there exists a function m̃ : (0, +∞) → (0, +∞) such
that lim"→0 m̃(") = 0 and the function u" − m̃(") is a viscosity subsolution of (3.2.15) in Q " .
Proof. The proof is similar to that of Lemma 3.2.10. The difference is that we assume that u is Lipschitz
continuous in Q. This is no longer a consequence of the assumption [H3].
The following theorem is a local Comparison Principle.
e 3]. Let u be a bounded, usc subsolution of (3.2.15) in
Theorem 3.3.3. Assume [H0], [H1], [H2] and [H
S and v be a bounded, lsc supersolution of (3.2.15) in S . Let R > 0 be as in (3.3.1). Let y0 ∈ Γ be fixed.
Then, if we set Q = B( y0 , R) ∩ S , we have
k (u − v)+ k L ∞ (Q) ≤k (u − v)+ k L ∞ (∂ Q) .
(3.3.53)
e 3] instead of [H3], we lose the lipschitz continuity of u in a neighborhood
Proof. Step 1 : By assuming [H
of Γ , which was an important property to prove Theorem 3.2.5. The first step consists therefore of
regularizing the subsolution so that it becomes Lipschitz continuous. Take α, p > 0 two positive numbers
and consider uα (= uα,p ) the sup-convolution of u with respect to the x 0 -variable defined in (3.3.17).
We chose α, p small enough so that Lemma 3.3.10 can be applied. Thus, from Lemma 3.3.10, we
know that uα is Lipschitz continuous in Q α and that there exists m : (0, +∞) → (0, +∞) such that
limα→0 m(α) = 0 and uα − m(α) is a subsolution of (3.2.15) in Q α . The definition of the set Q α (= Q α,p )
is given in (3.3.32).
Step 2 : We are now able to follow the proof of Theorem 3.2.5. The next step consists of a second
regularization of the subsolution u which this time produces a C 1 function in Γ . Let Q α," be the set
defined by
o
n
q
p
Q α," := x ∈ Q : d ist(x, ∂ Q) > α (2 k u k∞ +α 2 )2/p − α + " .
We consider the function uα," defined on Q α," by
0
0
uα," (x 0 e0 + x ) = uα ∗ ρ" (x 0 e0 + x ) =
Z
uα ((x 0 − τ)e0 + x 0 )ρ" (τ)dτ,
R
where ρ" is a sequence of mollifiers defined on R. It is clear that uα," is a C 1 function in Γ ∩ Q α," .
Besides, from Lemma 3.3.11, k uα," − uα k L ∞ (Q α," ) tends to 0 as " tends to 0 and there exists a function
m̃ : (0, +∞) → (0, +∞), such that lim"→0 m̃(") = 0 and such that the function uα," − m(α) − m̃(") is a
viscosity subsolution of (3.2.15) in Q α," .
Step 3 : Let us prove that
k (uα," − m(α) − m̃(") − v)+ k L ∞ (Q α," ) ≤k (uα," − m(α) − m̃(") − v)+ k L ∞ (∂ Q α," ) ,
(3.3.54)
for a fixed pair (α, ") of positive numbers.
Let Mα," be the supremum of uα," − m(α) − m̃(") − v on Q α," . The latter is reached for some x̄ α," ∈ Qα," ,
because the function uα," − m(α) − m̃(") − v is usc. If Mα," ≤ 0, then we clearly have k (uα," − m(α) −
m̃(") − v)+ k L ∞ (Q α," ) ≤k (uα," − m(α) − m̃(") − v)+ k L ∞ (∂ Q α," ) . So, we assume that Mα," > 0 and we want
to show that x̄ α," ∈ ∂ Q α," . Assume by contradiction that x̄ α," 6∈ ∂ Q α," . Then, x̄ α," is a local maximum
of uα," − m(α) − m̃(") − v.
1. If x̄ α," ∈
/ Γ : The usual doubling of variables method, with the auxiliary function ψβ (x, y) =
uα," (x) − m(α) − m̃(") − v( y) − d 2 (x, x̄ α," ) −
d 2 (x, y)
β2 ,
leads us to a contradiction.
2. If x̄ α," ∈ Γ : According to Lemma 3.3.10, uα," is Lipschitz continuous and C 1 with respect to
x 0 in Q̄ α," . Then, with a similar argument as in the proof of Lemma 3.2.7 we can construct a
test-function ϕ ∈ R(S ) such that ϕ|Γ = uα," |Γ and ϕ remains below uα," in a neighborhood of
Γ (take for example ϕ(x 0 e0 + x i ei ) = uα," (x 0 e0 ) − C x i with C great enough). It is easy to check
e 3],
that v − ϕ has a local minimum at x̄ α," . Then, we can use Theorem 3.2.4, which holds with [H
and we have two possible cases:
3.3. SECOND CASE : NORMAL CONTROLLABILITY NEAR INTERFACE
107
[B] λv(x̄ α," ) + HΓT (x̄ α," , D uα," |Γ (x̄ α," )) ≥ 0.
Moreover, uα," − m(α) − m̃(") is a subsolution of (3.2.15) which is C 1 on Γ . Then, according to Lemma 3.2.9, which can be apply here from
Remark 3.2.12, we have the inequality
λ(uα," (x̄ α," ) − m(α) − m̃(")) + HΓT (x̄ α," , D uα," |Γ (x̄ α," )) ≤ 0. Therefore, we obtain that
Mα," = uα," (x̄ α," ) − m(α) − m̃(") − v(x̄ α," ) ≤ 0, which is a contradiction.
[A] With the notations of Theorem 3.2.4, we have that
Z ηk
`i ( y x k (s), αki (s))e−λs ds + v( y x k (ηk ))e−ληk .
v(x k ) ≥
0
Moreover, from Lemma 3.2.9, which can be apply here from Remark 3.2.12,
Z ηk
`i ( y x k (s), αki (s))e−λs ds + (uα," ( y x k (ηk )) − m(α) − m̃("))e−ληk .
uα," (x k ) − m(α) − m̃(") ≤
0
Therefore
uα," (x k ) − m(α) − m̃(") − v(x k ) ≤ (uα," ( y x k (ηk )) − m(α) − m̃(") − v( y x k (ηk )))e−ληk .
Letting k tend to +∞, we find that Mα," ≤ Mα," e−λη , therefore that Mα," ≤ 0, which is a
contradiction.
Step 4 : In order to prove the final result, we have to pass to the limit as " tends to 0 and then as α
tends to 0. Let α > 0 be fixed. Let "0 be a strictly positive number and y be in Q α,"0 . Then, for all
0 < " < "0 we have that
(uα," ( y) − m(α) − m̃(") − v( y))+ ≤k (uα − m(α) − m̃(") − v)+ k L ∞ (∂ Q α," ) .
(3.3.55)
However, lim sup"→0 k (uα," − m(α) − m̃(") − v)+ k L ∞ (∂ Q α," ) ≤k (uα − m(α) − v)+ k L ∞ (∂ Q α ) . Indeed, the
supremum k (uα," − m(α) − m̃(") − v)+ k L ∞ (∂ Q α," ) is reached for some x α," in ∂ Q α," . Thus, for any
subsequence such that k (uα," − m(α) − m̃(") − v)+ k L ∞ (∂ Q ) converges to a limit ¯` when " tends to 0,
α,"
we can assume that x α," converges to some x̄ α that belongs to ∂ Q α when " tends to 0. Therefore, since
k uα," − uα k L ∞ (Q α," ) tends to 0 as " tends to 0, since uα is continuous in Q α and from the lower-semicontinuity of v, we have that ¯` ≤ (uα (x̄ α ) − m(α) − v(x̄ α ))+ ≤k (uα − m(α) − v)+ k L ∞ (∂ Q ) . Therefore,
α
by the pointwise convergence of uα," to uα , passing to the lim sup as " tends to 0 in (3.3.55) we deduce
(uα ( y) − m(α) − v( y))+ ≤k (uα − m(α) − v)+ k L ∞ (∂ Q α ) .
The above inequality is true for all y ∈ Q α,"0 , with "0 arbitrarily chosen, then
k (uα − m(α) − v)+ k L ∞ (Q α ) ≤k (uα − m(α) − v)+ k L ∞ (∂ Q α ) .
We are left with taking the limit as α tends to 0.
Fix now α0 and y ∈ Q α0 . For all 0 < α ≤ α0 we have
(uα ( y) − m(α) − v( y))+ ≤k (uα − m(α) − v)+ k L ∞ (∂ Q α ) .
(3.3.56)
As above, we have that lim supα→0 k (uα − m(α) − v)+ k L ∞ (∂ Q α ) ≤k (u − v)+ k L ∞ (∂ Q) . Indeed, the
supremum k (uα − m(α) − v)+ k L ∞ (∂ Q α ) is reached for some x α in ∂ Q α . Thus, for any subsequence
such that k (uα − m(α) − v)+ k L ∞ (∂ Q α ) converges to a limit ` as α tends to 0, we can assume that x α
converges to x̄ which belongs to ∂ Q when " tends to 0. But, from the properties of the sup-convolution,
the fact that u is upper-semi-continuous, continuous with respect to x 0 and the fact that v is lower-semicontinuous it is easy to check that necessarily ` ≤ (u(x̄) − v(x̄))+ ≤k (u − v)+ k L ∞ (∂ Q) . Therefore, by
the pointwise convergence of uα to u, passing to the lim sup as α tends to 0 in (3.3.56) we deduce
(u( y) − v( y))+ ≤k (u − v)+ k L ∞ (∂ Q) ,
∀ y ∈ Q α0 .
The above inequality is true ∀ y ∈ Q α0 , with α0 arbitrarily chosen, then
k (u − v)+ k L ∞ (Q) ≤k (u − v)+ k L ∞ (∂ Q) .
108
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
We are now able to prove the following global Comparison Principle.
e 3]. Let u be a bounded, usc subsolution of (3.2.15) in
Theorem 3.3.4. Assume [H0], [H1], [H2] and [H
S and v be a bounded, lsc supersolution of (3.2.15) in S . Then, u ≤ v in S .
Proof. The first step consists in a localization of the problem. For a some positive number K, we consider
p
M +M +1
the function ψ(x) := −K − 1 + |x|2 . It is easy to check that for K ≥ f λ l , ψ satisfies the viscosity
inequality
λψ + sup {−Dψ(x).ξ − ζ} ≤ −1.
(ξ,ζ)∈FL(x)
Then, if we set, for µ ∈ (0, 1), uµ = µu + (1 − µ)ψ, by convexity properties, we have that
λuµ +
sup
{−Duµ (x).ξ − ζ} ≤ −(1 − µ),
(ξ,ζ)∈FL(x)
where the above inequality is to be understood in the sense of the viscosity. In particular, uµ is a
subsolution of (3.2.15) in S . We set Mµ := sup x∈S {uµ (x) − v(x)}. Since uµ (x) is usc and tends to
−∞ as |x| tends to +∞ and since v is bounded, lsc the above supremum is reached at some x µ ∈ S .
We argue by contradiction, assuming that M := sup x∈S {u(x) − v(x)} > 0. Then, since Mµ tends to M
as µ tends to 1, for µ close enough to 1 we have Mµ > 0. We fix such a µ and we distinguish two cases.
1. If x µ ∈ Pi \ Γ for some i ∈ {1, . . . , N }, then a classical doubling variables method leads to a
contradiction.
2. If x µ ∈ Γ , then we are going to obtain a contradiction from Theorem 3.3.3. Let r > 0 be small
enough such that for all i ∈ {1, . . . , N } and x ∈ B(Γ , r) ∩ Pi
δ δ
[− , ] ⊂ { f i (x, a).ei : a ∈ Ai } .
2 2
We set Q µ := B(x µ , r) ∩ S and we consider the function ūµ defined in S by ūµ (x) = uµ (x) − |x −
x µ |2 (1 − µ)2 . It is easy to check that if µ is close enough to 1, ūµ is a subsolution of (3.2.15) in
S . Indeed, a direct computation gives
λūµ (x) +
sup
{−Dūµ (x, ξ) − ζ} ≤ −(1 − µ) + 2r M f (1 − µ)2 ,
(ξ,ζ)∈FL(x)
and the right hand side of this inequality is clearly negative if µ ∈ (0, 1) ∩ [1 − 2r1M f , 1). Then, we
apply Theorem 3.3.3 with Q = Q µ and the pair of sub/supersolution (uµ , v) : this leads to
Mµ = uµ (x µ ) − v(x µ ) = ūµ (x µ ) − v(x µ ) ≤k (ūµ − v)+ k L ∞ (∂ Q µ ) .
(3.3.57)
However, if x ∈ ∂ Q µ
ūµ (x) − v(x) = uµ (x) − v(x) − r 2 (1 − µ)2 ≤ Mµ − r 2 (1 − µ)2 < Mµ ,
in contradiction with (3.3.57).
Finally, we deduce that M ≤ 0 and the proof is complete.
As a consequence, we have the following result of uniqueness and regularity.
e 3]. Then, the value function v is continuous and is the
Theorem 3.3.5. Assume [H0], [H1], [H2] and [H
unique viscosity solution of (3.2.15) in S .
Proof. It is clear that Theorem 3.3.4 implies the uniqueness for the Hamilton-Jacobi equation (3.2.15).
We have just to prove that v is a solution of this equation. But, according to Theorem 3.3.2, v is a
discontinuous solution of (3.2.15) in S . From Theorem 3.3.4 applied to the pair of sub/supersolution
(v ? , v? ), we deduce v ? ≤ v? in S . Finally, v = v ? = v? and v is continuous.
3.4. EXTENSION TO A MORE GENERAL FRAMEWORK WITH ADDITIONAL DYNAMICS AND COST
AT THE INTERFACE
109
3.4
Extension to a more general framework with additional dynamics and cost at the interface
e 3], it is possible to extend all the results presented above to the case when there
With either [H3] or [H
e 3], we
are additional dynamics and cost at the interface. Since the framework is more general with [H
only discuss this case. We keep the setting from § 3.3 except that we take into account a set of controls
A0 , a dynamics f0 : Γ × A0 7−→ Re0 and a running cost `0 : Γ × A0 7−→ R. The assumptions made on A0 ,
f0 and `0 are the following.
(i) A0 is a non empty compact subset of the metric space A, disjoint from the other sets Ai , i ∈
{1, . . . , N }.
(ii) The function f0 satisfies the same boundedness and regularity properties as the functions f i ,
i ∈ {1, . . . , N }, described in [H0].
(ii) The function `0 satisfies the same boundedness and regularity properties as the functions `i ,
i ∈ {1, . . . , N }, described in [H1].
We define
M = (x, a) : x ∈ S , a ∈ Ai if x ∈ S \ Γ , and a ∈ ∪Ni=0 Ai if x ∈ Γ ,
the dynamics
∀(x, a) ∈ M ,
f (x, a) =
§
f i (x, a)
f i (x, a)
if x ∈ Pi \Γ , i ∈ {1, . . . , N }
if x ∈ Γ and a ∈ Ai , i ∈ {0, 1, . . . , N },
`(x, a) =
§
`i (x, a)
`i (x, a)
if x ∈ Pi \Γ , i ∈ {1, . . . , N }
if x ∈ Γ and a ∈ Ai , i ∈ {0, 1, . . . , N }.
and the running cost
∀(x, a) ∈ M ,
The infinite horizon optimal control problem is then given by (3.2.7) and (3.2.10). Then, we consider
the Hamilton-Jacobi equation (3.2.15) with the new definition of FL(x) :
§
FLi (x) S
if x belongs to Pi \Γ
FL(x) =
FL0 (x) ∪ i=1,...,N FL+i (x)
if x ∈ Γ ,
where for x ∈ Γ , FL0 (x) = {( f0 (x, a), `0 (x, a)) : a ∈ A0 }. The notion of viscosity sub and supersolutions
of (3.2.15) can be also defined as in (3.2.16) and (3.2.17). We obtain that the value function is discontinuous viscosity solution of (3.2.15) in S in the same manner as above, by passing by the relaxed
Hamilton-Jacobi equation (3.2.28). Note that the key result to pass from (3.2.15) to (3.2.28), Lemma
3.2.2, in the present framework becomes
e
f`(x)
=
e
f`(x) =
FL(x)
S
¦
Š©
S €
+
co
FL
(x)
∪
FL
(x)
∪
FL
(x)
∩
(Re
×
R)
0
j
0
j6=i
i=1,...,N
i
if x ∈ S \Γ ,
if x ∈ Γ .
The proof of this can be made in the same way as above.
As previously, we have some equivalent definitions for (3.2.16) and (3.2.17) given by :
• An upper semi-continuous function u : S → R is a subsolution of (3.2.15) in S if for any x ∈ S ,
any ϕ ∈ R(S ) s.t. u − ϕ has a local maximum point at x, then
λu(x) + H i (x, D ϕ|Pi (x)) ≤ 0 if x ∈ Pi \Γ ,
λu(x) + HΓ x, D (ϕ|Γ ) (x), D ϕ|P1 (x), . . . , D ϕ|PN (x) ≤ 0
if x ∈ Γ .
• A lower semi-continuous function u : S → R is a supersolution of (3.2.15) if for any x ∈ S , any
ϕ ∈ R(S ) s.t. u − ϕ has a local minimum point at x, then
λu(x) + H i (x, D ϕ|Pi (x)) ≥ 0 if x ∈ Pi \Γ ,
λu(x) + HΓ x, D (ϕ|Γ ) (x), D ϕ|P1 (x), . . . , D ϕ|PN (x) ≥ 0
if x ∈ Γ .
110
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Where the new definition of HΓ : Γ × Re0 ×
Q
i=1,...,N (Re0
× Rei ) → R is given by
§
ª
HΓ (x, p0 , p1 , . . . , pN ) = max H0 (x, p0 ) max H i+ (x, pi ) ,
i=1,...,N
where the Hamiltonians H i+ are defined in (3.2.20) and the Hamiltonian H0 : Γ × Re0 7−→ R, is defined
by
H0 (x, p0 ) = max (− f0 (x, a).p0 − `0 (x, a)) .
a∈A0
The tangential Hamiltonian at Γ , HΓT : Γ × Re0 7−→ R is also slightly changed,
§
ª
HΓT (x, p) = max H0 (x, p), max HΓT,i (x, p) ,
i=1,...,N
where the Hamiltonians HΓT,i are defined in (3.2.22).
With these new definitions, all the results proved in § 3.3 hold with obvious modifications of the proofs.
In particular,
• a subsolution of the present problem is also a subsolution of the former problem. So Lemma 3.2.7
(with remark 3.2.8) and Lemma 3.2.9 (with remark 3.2.12) hold.
• Lemma 3.2.1 holds since it only involves the Hamiltonians H i , H i+ and HΓT,i .
• The proof of Theorem 3.2.4 can still be used. In particular, with the choice of (qi )i=1,...,N made in
this proof, we have the identity
HΓ (x̄, , D(φ|Γ )(x̄) + q1 e1 , . . . , D(φ|PN )(x̄) + qN eN ) = HΓT (x̄, D(φ|Γ )(x̄)).
• The proofs of the regularisation results, Lemma 3.3.10 and Lemma 3.3.11, are unchanged.
• The proofs of the Comparison principles, Theorem 3.3.3 and Theorem 3.3.4 , are unchanged.
3.5
3.5.1
Appendix
Proof of the Proposition 3.2.1
The proof of the dynamic programming principle, Proposition 3.2.1, is essentially based on the following
semi-group property.
Proposition 3.5.1. For all ( y, α) in T x and for all t > 0, we have that ( y( . + t), α( . + t)) belongs to T y(t) .
Proof. It is a simple consequence of the equality
y(s + t) = y(t) +
Z
s
f ( y(u + t), α(u + t))du.
(3.5.1)
0
Now prove the Proposition 3.2.1.
Proof. Let t > 0 be fixed. We consider the function w defined on S by
w(x) =
We wish to prove that v = w.
Z
inf
( y x ,α)∈T x
0
t
`( y x (s), α(s))e
−λs
ds + v( y x (t))e
−λt
.
3.5. APPENDIX
111
• First, show that v ≥ w : Let x be in S and ( y x , α) be in T x . Then, we have
J(x, ( y x , α))
=
=
Rt
R +∞
`( y x (s), α(s))e−λs ds + t
`( y x (s), α(s))e−λs ds
0
R
Rt
+∞
`( y x (s), α(s))e−λs ds + e−λt 0 `( y x (t + s), α(t + s))e−λs ds.
0
But according to (3.5.1), ( y( . + t), α( . + t)) ∈ T y x (t) and therefore
J(x, ( y x , α))
=
≥
Rt
`( y x (s), α(s))e−λs ds + e−λt J( y x (t), ( y x ( . + t), α( . + t))
R0t
`( y x (s), α(s))e−λs ds + e−λt v( y x (t)).
0
The last inequality is true for any ( y x , α) ∈ T x , therefore taking the infimum over ( y x , α) in T x
we get v ≥ w.
• Secondly, show that w ≥ v : Let x be in S and ( y x , α) be in T x . From the definition of v( y x (t)),
we have the existence of ( y1 , α1 ) in T y x (t) such that
v( y x (t)) ≥ J( y x (t), ( y1 , α1 )) − ".
(3.5.2)
Let ( ȳ, ᾱ) be the admissible controlled trajectory belongs to T x defined as follows
ᾱ(s) =
§
α(s)
α1 (s − t)
if
if
s ≤ t,
s > t;
ȳ(s) =
§
y x (s)
y1 (s − t)
if
if
s ≤ t,
s > t.
and
It is easy to check that ( ȳ, ᾱ) belongs to T x . Then, we have
v(x)
≤
=
=
=
=
J(x, ( ȳ, ᾱ))
Rt
R +∞
`( ȳ(s), ᾱ(s))e−λs ds + t
`( ȳ(s), ᾱ(s))e−λs ds
0
Rt
R
+∞
`( ȳ(s), ᾱ(s))e−λs ds + e−λt 0 `( ȳ(t + s), ᾱ(t + s))e−λs ds
R0t
R
+∞
`( y x (s), α(s))e−λs ds + e−λt 0 `( y1 (s), α1 (s))e−λs ds
R0t
`( y x (s), α(s))e−λs ds + e−λt J( y x (t), ( y1 , α1 )).
0
Then, according to (3.5.2)
Z
v(x) ≤
t
`( y x (s), α(s))e−λs ds + e−λt (v( y x (t)) + ").
(3.5.3)
0
Like the inequality (3.5.3) is true for all " > 0, we have
Z
v(x) ≤
t
`( y x (s), α(s))e−λs ds + e−λt v( y x (t)),
0
and taking the infimum over ( y x , α) in T x we get w ≥ v.
3.5.2
Proof of the Proposition 3.2.2
The proof of this result is essentially a consequence of the following Lemma.
Lemma 3.5.1. There exists r > 0 and C > 0 such that for all x, z in S ∩ B(Γ , r), there exists ( y x,z , α x,z )
in T x and τ x,z ≥ 0 such that y x,z (τ x,z ) = z and τ x,z ≤ C d(x, z). Where we recall that B(Γ , r) = {z ∈ Rd :
dist(z, Γ ) < r} and that d(. , .) is the geodesic distance on S .
Now, prove the proposition 3.2.2.
112
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Proof. Let x be in S . We want to prove that v is continuous at x.
1st step: Show that lim supz→x v(z) ≤ v(x). Let " be a strict positive number. By definition of v(x),
there exists ( ȳ x , ᾱ) ∈ T x such that
J(x, ( ȳ x , ᾱ)) < v(x) + ".
(3.5.4)
1. If x ∈ B(Γ , r) ∩ S : Then, for z ∈ B(Γ , r) ∩ S , we introduce ( y z , αz ) the admissible controlled
trajectory of Tz defined as follows :
§
αz,x (t)
if t ≤ τz,x ,
αz (t) =
ᾱ(t − τz,x ) if t > τz,x ;
and
y (t) =
z
yz,x (t)
ȳ x (t − τz,x )
§
if
if
t ≤ τz,x ,
t > τz,x .
Where ( yz,x , αz,x ) is given by Lemma 3.5.1 and ( ȳ x , ᾱ) is the admissible controlled trajectory of
T x from (3.5.4). It is easy to check that ( y z , αz ) belongs to Tz . We have
v(z)
≤
=
≤
J(z, ( y z , αz ))
R τz,x
R +∞
`( yz,x (t), αz,x (t))e−λt d t + e−λτz,x 0 `( ȳ x (t), ᾱ(t))e−λt d t
0
M` τz,x + e−λτz,x J(x, ( ȳ x , ᾱ)).
Then, according to Lemma 3.5.1 and the inequality (3.5.4), we have
M` C d(z, x) + v(x) + ".
≤
v(z)
(3.5.5)
In (3.5.5) we can take any z in B(Γ , r) ∩ S and we can assume " as small as desired. Then, this
inequality gives us lim supz→x v(z) ≤ v(x).
2. If x ∈ Pi \ B(Γ , r) for some i ∈ {1, . . . , N } : Then, like we want take the limite as z tends to
x, we can just consider z in Pi \ Γ . Let ȳz : [0, T̄ ) → S be the unique maximal solution of
Rt
ȳz (t) = z + 0 f i ( ȳz (s), ᾱ(s))ds staying in S , where ᾱ is the control law of the inequality (3.5.4).
Since x ∈ Pi \ B(Γ , r), from [H0] T̄ > 0. We introduce
Tz = inf{t ∈ [0, T̄ ) : ȳz (t) ∈ Γ }
andTx,z = inf{t ∈ [0, T̄ ) : ȳ x (t) ∈ Γ or ȳz (t) ∈ Γ },
where we use the convention that Tz = +∞ (resp. Tx,z = +∞) if {t ∈ [0, T̄ ) : ȳz (t) ∈ Γ } (resp.
{t ∈ [0, T̄ ) : ȳ x (t) ∈ Γ or ȳz (t) ∈ Γ }) is empty. Since x and z belong to Pi \ Γ and by definition of
Tz and Tx,z , we know that T̄ ≥ T̄z ≥ Tx,z > 0. We introduce also the new notations
T" = −
"λ
r e LT"
1
log(
) and ρ" =
.
λ
2Ml
4
Now, we assume that |z − x| ≤ ρ" .
(a) If Tx,z = +∞ : Then T̄ = +∞ and ( ȳz , ᾱ) belongs to Tz . So
v(z)
≤
≤
J(z, ( ȳz , ᾱ))
|J(z, ( ȳz , ᾱ)) − J(x, ( ȳ x , ᾱ))| + J(x, ( ȳ x , ᾱ)).
However
|J(z, ( ȳz , ᾱ)) − J(x, ( ȳ x , ᾱ))|
≤
≤
=
R T"
| 0 (`( ȳz (t), ᾱ(t)) − `( ȳ x (t), ᾱ(t))) e−λt d t|
R +∞
+ T (`( ȳz (t), ᾱ(t)) − `( ȳ x (t), ᾱ(t))) e−λt d t|
R T" "
R +∞
| 0 ω` (| ȳz (t) − ȳ x (t)|)e−λt d t| + 2M` T e−λt d t
"
R T"
| 0 ω` (| ȳz (t) − ȳ x (t)|)e−λt d t| + ".
But, since ȳ x and ȳz have the same control law ᾱ, according to Gronwall Lemma, we have
| ȳz (t) − ȳ x (t)| ≤ |x − z|e L f t ,
3.5. APPENDIX
113
and then
|J(z, ( ȳz , α)) − J(x, ( ȳ x , ᾱ))|
≤
ω` (|x − z|e L f T" )T" + ".
(3.5.6)
Finally, according to (3.5.4) and (3.5.6) we obtain
v(z) ≤ v(x) + ω` (|x − z|e L f T" )T" + 2".
First taking the limsup as z tends to x and then the limite as " tends to 0 we obtain
lim supz→x v(z) ≤ v(x).
(b) If Tx,z < +∞ : Since T" tends to +∞ as " tends to 0, we can assume that Tx,z < T" . Then,
since we take z such that |x − z| ≤ ρ" and since ȳ x (Tx,z ) or ȳz (Tx,z ) belongs to Γ , a simple
calcuation show that ȳ x (Tx,z ) and ȳz (Tx,z ) belong to B(Γ , r). Therefore, we can consider
( y ȳz (Tx,z ), ȳ x (Tx,z ) , α ȳz (Tx,z ), ȳ x (Tx,z ) ) ∈ T ȳz (Tx,z ) and τ ȳz (Tx,z ), ȳ x (Tx,z ) ≥ 0, being respectively the path
which connects ȳz (Tx,z ) to ȳ x (Tx,z ) and the time to reach ȳ x (Tx,z ) from ȳz (Tx,z ) given by
Lemma 3.5.1. We introduce ( ỹ, α̃) the admissible controlled trajectory of Tz defined as
follows :

if t < Tx,z ,
 ᾱ(t)
α
(t
−
T
)
if Tx,z ≤ t < Tx,z + τ ȳz (Tx,z ), ȳ x (Tx,z ) ,
(3.5.7)
α̃(t) =
ȳz (Tx,z ), ȳ x (Tx,z )
x,z
 ᾱ(t − τ
if Tx,z + τ ȳz (Tx,z ), ȳ x (Tx,z ) ≤ t;
ȳz (Tx,z ), ȳ x (Tx,z ) )
and

 ȳz (t)
y ȳz (Tx,z ), ȳ x (Tx,z ) (t − Tx,z )
ỹ(t) =
 ȳ (t − τ )
x
z,x
if
if
if
t < Tx,z ,
T ≤ t < Tx,z + τ ȳz (Tx,z ), ȳ x (Tx,z ) ,
Tx,z + τ ȳz (Tx,z ), ȳ x (Tx,z ) ≤ t.
(3.5.8)
Where ( ȳ x , ᾱ) is the admissible controlled trajectory of T x from (3.5.4). We have
v(z)
≤
=
J(z, ( ỹ, α̃))
R Tx,z
R Tx,z +τ ȳ (T ), ȳ (T )
z x,z
x x,z
`( ỹ(t), α̃(t))e−λt d t + T
`( ỹ(t), α̃(t))e−λt d t
0
x,z
R +∞
+ T +τ
`( ỹ(t), α̃(t))e−λt d t.
x,z
ȳz (T x,z ), ȳ x (T x,z )
By construction, we have
R +∞
Tx,z +τ ȳz (T x,z ), ȳ x (T x,z )
`( ỹ(t), α̃(t))e−λt d t
≤
J(x, ( ȳ x , ᾱ)) −
R Tx,z
0
`( ȳ x (t), ᾱ(t))e−λt d t,
and therefore
R Tx,z
R Tx,z +τ ȳ (T ), ȳ (T )
z x,z
x x,z
v(z) ≤
(`( ỹ(t), α̃(t)) − `( ȳ(t), ᾱ(t))) e−λt d t + T
`( ỹ(t), α̃(t))e−λt d t
0
+J(x, ( ȳ x , ᾱ)).
x,z
Then, calculations similar to those of the previous case 2a and (3.5.4) lead us to
v(z) ≤ Tx,z ω` (e Tx,z L |x − z|) + M` C e LTx,z |x − z| + v(x) + ",
which implies that lim supz→x v(z) ≤ v(x).
2nd step: Show that lim infz→x v(z) ≥ v(x). The first step, consist to show that for z close enough to x,
we have
v(z) ≤ v(x) + ω" (d(x, z)) + 2",
(3.5.9)
where ω" (.) is a modulus of continuity which depends on " and d(., .) is the geodesic distance on S .
But, it is easy to check that by changing the roles of x and z in the calculation of the first step above we
get exactly (3.5.9) for z enough near to x.
Therefore, taking the liminf as z tends to x and then the limite as " tends to 0 in (3.5.9) we obtain
lim infz→x v(z) ≥ v(x).
114
3.5.3
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
Proof of the Proposition 3.2.3
Proof. According to the proposition 3.2.2, we have already that v is continuous. Then, we just need to
check that v satisfies the sub and supersolution inequalities.
The subsolution inequality : Let x be in S and ϕ be in R(S ) such that v − ϕ has a local maximum
at x. We want to show that
λv(x) +
{−Dϕ(x, ξ) − ζ} ≤ 0.
sup
(3.5.10)
(ξ,ζ)∈ f˜l(x)
Since v − ϕ has a local maximum at x, there exists r > 0 such that ∀z ∈ B(x, r) ∩ S ,
ϕ(x) − ϕ(z) ≤ v(x) − v(z).
(3.5.11)
On the other hand, for (ξ, ζ) ∈ f˜l(x), there exists ( y x,n , αn ) and t n → 0+ such that
ξ
=
ζ
=
1
tn
limn→+∞ t1n
limn→+∞
R tn
R0t n
0
f ( y x,n (t), αn (t))d t = limn→+∞
y x,n (t n )−x
,
tn
`( y x,n (t), αn (t))d t.
According to Assumption [H0], there exists T > 0 such that for all ( y x , α) ∈ T x , y x (t) ∈ B(x, r) ∩ S
for t < T . We can assume that for all n ∈ N, t n ∈ [0, T ]. Then, according to (3.5.11) and the dynamic
programming principle, proposition 3.2.1, for all n ∈ N we have
ϕ(x) − ϕ( y x,n (t n ))
≤
≤
v(x) − v( y x,n (t n ))
R tn
−λt
−λt n
`(
y
(t),
α
(t))e
d
t
+
v(
y
(t
))
e
−
1
.
x,n
n
x,n
n
0
(3.5.12)
Then, the idea is to divide by t n and to take the limite as n tends to +∞. We obtain easily that
lim
n→+∞
1
tn
Z
tn
`( y x,n (t), αn (t))e
−λt
0
e−λt n − 1
d t + v( y x,n (t n ))
tn
= ζ − λv(x).
Otherwise, since y x,n (t n ) = x + t n (ξ + ◦(1)), if we show that ξ + ◦(1) ∈ Tx (S ) we get that
lim
ϕ(x) − ϕ( y x,n (t n ))
tn
n→+∞
= −Dϕ(x, ξ).
(3.5.13)
If x ∈ Pi \ Γ , for some i ∈ {1, . . . , N }, if n is large enough, we have that y x,n (t n ) belongs to Pi \ Γ and
that ξ+◦(1) belongs to Pi = Tx (S ). Then, if x ∈ Γ , ξ+◦(1) =
according to (3.5.13) and (3.5.12) we get
y x,n (t n )−x
tn
belongs to S ⊂ Tx (S ). Finally,
λv(x) − Dϕ(x, ξ) − ζ ≤ 0,
and therefore (3.5.10) is true.
Show the supersolution inequality : Let x be in S and ϕ be in R(S ) such that v − ϕ has a local
minimum at x. We want to show that
λv(x) +
sup
{−Dϕ(x, ξ) − ζ} ≥ 0.
(3.5.14)
(ξ,ζ)∈ f˜l(x)
Since v − ϕ has a local minimum at x, there exists r > 0 such that ∀z ∈ B(x, r) ∩ S ,
ϕ(x) − ϕ(z) ≥ v(x) − v(z).
(3.5.15)
From the dynamic programming principle, Proposition 3.2.1, for " > 0 and t > 0, there exists ( y x , α) ∈
T x (depending on " and t) such that
v(x) + t"
≥
≥
Rt
`( y x (s), α(s))e−λs ds + e−λt v( y x (t))
R0t
`( y x (s), α(s))ds + e−λt v( y x (t)) + ◦ t→0 (t),
0
(3.5.16)
3.5. APPENDIX
115
where we use the boundedness of ` to get he last line. Then, according to (3.5.15), for t sufficiently
small, we get
Z t
`( y x (s), α(s))ds + (1 − eλt )v( y x (t)) ≥ −t" + ◦ t→0 (t).
ϕ(x) − ϕ( y x (t)) −
(3.5.17)
0
Let t n be a sequence such that t n → 0+ . Let ( y x,n , αn ) be in T x associated with " and t n , so that (3.5.16)
are true. We assume that t n small enough so that (3.5.17) are true. Then, we have
Z tn
`( y x,n (s), αn (s))ds + (1 − eλt n )v( y x,n (t n )) ≥ −t n " + ◦(t n ).
ϕ(x) − ϕ( y x,n (t n )) −
(3.5.18)
0
x,n (t n )−x
R tn
Š
`(
y
(s),
α
(s))ds
is bounded in Tx (S ) × R, then we can assume that
x,n
n
tn
0
e
e
it converges to (ξ̄, ζ̄) as n tends to +∞. By definition of f`(x),
we clearly have (ξ̄, ζ̄) ∈ f`(x).
Finally,
since previously, dividing by t n in (3.5.18) and letting n tends to +∞. For the same reasons as above,
we obtain
λv(x) − Dϕ(x, ξ) − ζ ≥ 0,
The sequence
€y
,
1
tn
and therefore (3.5.14) is true.
3.5.4
Proof of Lemma 3.2.2
e
Proof. The proof of the equality¦ f`(x)
= FL(x)€for x ∈ S \Γ is standard
(see [BCD97], Lemma 2.41,
Š©
S
+
e
page 129), and the inclusions co FLi (x) ∪ j6=i FL j (x) ∩ (Γ × R) ⊂ f`(x) for x ∈ Γ and i ∈ {1, . . . , N }
are proved by explicitly constructing trajectories, see [ACCT13]. We skip this part. This leads to
S
FL(x)
¦
Š© =
S €
+
co
FL
(x)
∪
FL
(x)
∩
(Γ
×
R)
⊂
j
j6=i
i=1,...,N
i
e
f`(x)
e
f`(x)
if x ∈ S \Γ ,
if x ∈ Γ .
e
We now prove the reverse inclusion. Let x ∈ Γ . For any (ζ, µ) ∈ f`(x),
there exists a sequence of
admissible trajectories ( yn , αn ) ∈ T x and a sequence of times t n → 0+ such that
Z tn
Z tn
1
1
f ( yn (t), αn (t))d t = ζ, and lim
`( yn (t), αn (t))d t = µ.
lim
n→∞ t
n→∞ t
n 0
n 0
First, remark that by cronstruction ζ necessarily belongs to Re0 × R+ ei , for some i ∈ {1, . . . , N }.
• If ζ ∈
/ Re0 , then there exists an index i in {1, . . . , N } such that ζ ∈ (Re0 × R+ ei )\Re0 : in this case,
yn (t n ) ∈ Pi \Γ . Hence,
yn (t n ) = x +
f j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t +
0
j=1
with
tn
N Z
X
Z
Z
tn
f ( yn (t), αn (t))1 yn (t)∈Γ d t,
(3.5.19)
0
tn
Z0t n
f j ( yn (t), αn (t))1 yn (t)∈P j \Γ .e j d t = 0
if j 6= i,
(3.5.20)
f i ( yn (t), αn (t))1 yn (t)∈Pi \Γ .ei d t = yn (t n ).ei .
0
These identities are a consequence of Stampacchia’s theorem: consider for example j ∈ {1, . . . , N }
and the function κ j : y 7→ y1 y∈P j \Γ .e j . It is easy to check that t 7→ κ j ( yn (t)) belongs to
W 1,∞ (0, t n ) and that its weak derivative coincides almost everywhere with
t 7→ f j ( yn (t), αn (t))1 yn (t)∈P j \Γ .e j . This implies (3.5.20).
For j = 1, . . . , N , let t j,n be defined by
¦
©
t j,n = t ∈ [0, t n ] : yn (t) ∈ P j \Γ .
116
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
If j 6= i and t j,n > 0 then
1
tn
Z
f j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t,
t j,n
0
1
Z
=
t j,n
Z
tn
` j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t
0
tn
f j (x, αn (t))1 yn (t)∈P j \Γ d t,
Z
0
tn
` j (x, αn (t))1 yn (t)∈P j \Γ d t + o(1)
0
where
Š of
€R t o(1) is a vector tending to 0 asR nt → ∞. Therefore, the distance
n
n
1
f j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t, 0 ` j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t to the set FL j (x) tends to
t j,n
0
R tn
0. Moreover, according to (3.5.20), we have that 0 f j ( yn (t), αn (t))1 yn (t)∈P j \Γ .e j d t = 0. Hence,
€ Rt
Š
R tn
n
the distance of t 1j,n e j 0 f j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t, 0 ` j ( yn (t), αn (t))1 yn (t)∈P j \Γ d t to the set
€
Š
FL j (x) ∩ (Re0 × R) tends to zero as n tends to ∞.
If the set {t : yn (t) ∈ Γ } has a nonzero measure, then
1
|{t : yn (t) ∈ Γ }|
=
tn
Z
1
|{t : yn (t) ∈ Γ }|
f ( yn (t), αn (t))1 yn (t)∈Γ d t,
Z
0
`( yn (t), αn (t))1 yn (t)∈Γ d t
0
tn
Z
tn
f (x, αn (t))1 yn (t)∈Γ d t,
Z
0
tn
`(x, αn (t))1 yn (t)∈Γ d t + o(1)
0
€R t
Š
R tn
n
1
Therefore, the distance of |{t: yn (t)∈Γ
f
(
y
(t),
α
(t))1
d
t,
`(
y
(t),
α
(t))1
d
t
n
n
y
(t)∈Γ
n
n
y
(t)∈Γ
}|
n
n
0
0
©
¦S
N
to the set co
FL
(x)
tends
to
zero
as
n
tends
to
∞.
Moreover,
from theorem 3.2.2,
j
j=1
f ( yn (t), αn€(t)) ∈ Re0 almost everywhere on {t : yn (t) ∈ Γ }. Therefore,
the
distance
of
Š
R tn
R tn
1
f ( yn (t), αn (t))1 yn (t)∈Γ d t, 0 `( yn (t), αn (t))1 yn (t)∈Γ d t to the set
|{t: yn (t)∈Γ }|
¦S € 0
Š©
N
co
FL
(x)
∩
(Re
×
R)
tends to zero as n tends to ∞.
j
0
j=1
Finally, we know that Ti,n > 0.
1
=
tn
Z
f i ( yn (t), αn (t))1 yn (t)∈Pi \Γ d t,
t i,n
0
1
Z
t i,n
tn
`i ( yn (t), αn (t))1 yn (t)∈Pi \Γ d t
0
tn
f i (x, αn (t))1 yn (t)∈Pi \Γ d t,
0
Z
Z
tn
`i (x, αn (t))1 yn (t)∈Pi \Γ d t + o(1)
0
so the
€R tdistance of
Š
R tn
n
1
f i ( yn (t), αn (t))1 yn (t)∈Pi \Γ d t, 0 `i ( yn (t), αn (t))1 yn (t)∈Pi \Γ d t to the set FL+i (x) tends to
t i,n
0
zero as n tends to ∞.
Combining
all the observations above, we see that
of
€ Rt
Š the distance
¦
Š©
R tn
S €
n
1
1
+
f
(
y
(t),
α
(t))d
t,
`(
y
(t),
α
(t))d
t
to
co
FL
(x)
∪
FL
(x)
∩
(Re
×
R)
n
n
n
n
j
0
j6=i
tn 0
tn 0
i
¦
Š©
S €
tends to 0 as n → ∞. Therefore (ζ, µ) ∈ co FL+i (x) ∪ j6=i FL j (x) ∩ (Re0 × R) .
• If ζ ∈ Re0 , either there exists i such that yn (t n ) ∈ Pi \Γ or yn (t n ) ∈ Γ :
• If yn (t n ) ¦
∈ Pi \Γ , then we
can make exactlyŠ©
the same argument as above and conclude that
S €
(ζ, µ) ∈ co FL+i (x) ∪ j6=i FL j (x) ∩ (Re0 × R) . Since ζ ∈ Re0 , we have in fact that (ζ, µ) ∈
Š
SN €
co j=1 FL j (x) ∩ (Re0 × R) .
• if y (t ) ∈ Γ , according to Stampacchia theorem, we have that
Z tn n n
f j ( yn (t), αn (t))1 yn (t)∈P j \Γ .e j d t = 0 for all j = 1, . . . , N . We can repeat the argument above,
¦S €
Š©
N
and obtain that (ζ, µ) ∈ co
FL
(x)
∩
(Re
×
R)
.
j
0
j=1
0
3.5. APPENDIX
3.5.5
117
Proof of Theorem 3.3.2
Proof. First, remark that in this proof, the notation o" (1) will denote an application independent of
t, which tends to 0 as " tends to 0 and that for k ∈ N? the notation O(t k ) will denote an application
O(t k )
independent of ", such that t k remains bounded as t tends to 0. Show that v? is a supersolution
of (3.2.15) : for any x ∈ S , let (x " )">0 be a sequence such that x " tends to x when " tends to 0 and
v(x " ) tends to v? (x) when " tends to 0. Let ϕ be in R(S ) such that v? − ϕ has a local minimum at x,
i.e. there exists r > 0 such that
v? (x) − ϕ(x) ≤ v? ( y) − ϕ( y).
∀ y ∈ B(x, r) ∩ S ,
(3.5.21)
From the dynamic programming principle (Proposition 3.2.1), for any " > 0 and t > 0, there exists
( ȳ",t , ᾱ",t ) ∈ T x " such that
v(x " )
Rt
`( ȳ",t (s), ᾱ",t (s))e−λs ds + e−λt v( ȳ",t (t)) − "
R0t
`( ȳ",t (s), ᾱ",t (s))e−λs ds + e−λt v? ( ȳ",t (t)) − ".
0
≥
≥
Then, according to (3.5.21), for " and t > 0 small enough we have
v(x " ) − v? (x) ≥
Z
t
`( ȳ",t (s), ᾱ",t (s))e−λs ds + v? (x)(e−λt − 1) + (ϕ( ȳ",t (t)) − ϕ(x))e−λt − ".
0
Rt
Rt
Using that v(x " ) − v? (x) = o" (1), that 0 `( ȳ",t (s), ᾱ",t (s))e−λs ds = 0 `( ȳ",t (s), ᾱ",t (s))ds + O(t 2 ) and
that (ϕ( ȳ",t (t)) − ϕ(x))e−λt = (ϕ( ȳ",t (t)) − ϕ(x)) + t o" (1) + O(t 2 ), we finally obtain
Z
0≥
t
`( ȳ",t (s), ᾱ",t (s))ds + v? (x)(e−λt − 1) + ϕ( ȳ",t (t)) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1). (3.5.22)
0
• If x ∈ Pi \ Γ : since x " tends to x belonging to Pi \ Γ as " tends to 0 and since the dynamic f is
bounded, see remark 3.2.2, there exists t̄ > 0 such that for any t ∈ (0, t̄), ȳ",t (s) ∈ Pi \ Γ for any
s ∈ (0, t). Then, the inequality (3.5.22) can be rewritten as follows
t
Z
0
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )( ȳ",t (s)). f i ( ȳ",t (s), ᾱ",t (s))ds + v? (x)(e−λt − 1)
≥
0
+ϕ(x " ) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1).
Using that ϕ(x " ) − ϕ(x) = o" (1) and that D(ϕ|Pi )( ȳ",t (t)) = D(ϕ|Pi )(x) + o" (1) + O(t), we get
that
Z t
0
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )(x). f i ( ȳ",t (s), ᾱ",t (s))ds + v? (x)(e−λt − 1)
≥
0
(3.5.23)
2
+t o" (1) + O(t ) + o" (1).
€R t
Š
Rt
It is easy to check that 1t 0 f i ( ȳ",t (s), ᾱ",t (s))ds, 0 `i ( ȳ",t (s), ᾱ",t (s))ds is at a distance to FLi (x)
of the order of o" (1) + O(t). Thus,

‹
0 ≥ v? (x)(e−λt − 1) − t
max {−D(ϕ|Pi )(x, ξ) − ζ} + t o" (1) + O(t 2 ) + o" (1).
(ξ,ζ)∈FLi (x)
Finally, dividing the latter inequality by t, taking the limit as " tends to 0 and in a second time
the limit as t tends to 0 we get the desired inequality
λv? (x) +
max
{−D(ϕ|Pi )(x, ξ) − ζ} ≥ 0.
(ξ,ζ)∈FLi (x)
• If x ∈ Γ and x " ∈ Pi \ Γ : Let τ",t > 0 be the exit time of ȳ",t from Pi \ Γ . Up to the extraction of
a subsequence, we may assume that either τ",t ≤ t for all " > 0 or that τ",t > t for all ".
118
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
If τ",t > t : The same calculations as in the case where x ∈ Pi \ Γ give us (3.5.23). As above
Š
€R t
Rt
1
f ( ȳ",t (s), ᾱ",t (s))ds, 0 `i ( ȳ",t (s), ᾱ",t (s))ds is at a distance to FLi (x) of the order of o" (1)+
t
0 i
Rt
O(t). But this time, we have that 0 f i ( ȳ",t (s), ᾱ",t (s))ds.ei ≥ −x " .ei = o" (1) for all t, " > 0 and
then we have the more specific information that
Š
€R t
Rt
o (1)
1
f
(
ȳ
(s),
ᾱ
(s))ds,
`
(
ȳ
(s),
ᾱ
(s))ds
is at a distance to FL+i (x) of the order of "
+
",t
",t
",t
",t
t
0 i
0 i
t
o" (1) + O(t). Finally, (3.5.23) give us as desired
λv? (x) + max
max
{−D(ϕ|Pi )(x, ξ) − ζ} ≥ 0.
i∈{1,...,N } (ξ,ζ)∈FL+
i (x)
If τ",t ≤ t : Then, the inequality (3.5.22) can be written as follow
Z
≥
0
τ",t
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )( ȳ",t (s)). f i ( ȳ",t (s), ᾱ",t (s))ds
N Z t
X
”
—
+
1{ ȳ",t (s)∈P j \Γ } ` j ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|P j )( ȳ",t (s)). f j ( ȳ",t (s), ᾱ",t (s)) ds
0
+
Zj=1t
τ",t
τ",t
1{ ȳ",t (s)∈Γ } `( ȳ",t (s), ᾱ",t (s)) + D (ϕ|Γ ) ( ȳ",t (s)). f ( ȳ",t (s), ᾱ",t (s)) ds
+v? (x)(e−λt − 1)+ϕ(x " ) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1).
(3.5.24)
Note that to obtain the third line of this inequality, we use Point 3 of Theorem 3.3.1. To obtain the
supersolution inequality, we have to deal with each term of the inequality (3.5.24) individually.
Z t
→ Let j ∈ {1, . . . , N } be such that ȳ",t (t) 6∈ P j \Γ . Then
and
τ",t
1{ ȳ",t (s)∈P j \Γ } f j ( ȳ",t (s), ᾱ",t (s)).e j ds = 0
Œ
‚Z t
Z t
1
1{ ȳ",t (s)∈P j \Γ } ` j ( ȳ",t (s), ᾱ",t (s))ds
1{ ȳ",t (s)∈P j \Γ } f j ( ȳ",t (s), ᾱ",t (s))ds,
|{s : ȳ",t (s) ∈ P j \Γ }|
τ",t
τ",t
is at a distance to FLi (x) ∩ (Re0 × R) of the order of o" (1) + O(t). Thus, we get
Z
t
”
—
1{ ȳ",t (s)∈P j \Γ } ` j ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|P j )( ȳ",t (s)). f j ( ȳ",t (s), ᾱ",t (s)) ds
τ",t
€
Š
≥ |{s : ȳ",t (s) ∈ P j \Γ }| − max(ξ,ζ)∈FL j (x)∩(Re0 ×R) {−Dϕ(x, ξ) − ζ} + o" (1) + O(t) .
(3.5.25)
→ If there exists one k ∈ {1, . . . , N }, such that ȳ",t (t) ∈ Pk \ Γ . In this case, we have that
Z t
τ",t
1{ ȳ",t (s)∈Pk \Γ } f k ( ȳ",t (s), ᾱ",t (s)).ek ds = ȳ",t (t).ek > 0 and then that
1
|{s : ȳ",t (s) ∈ Pk \Γ }|
‚Z
t
τ",t
1{ ȳ",t (s)∈Pk \Γ } f k ( ȳ",t (s), ᾱ",t (s))ds,
Z
t
τ",t
Œ
1{ ȳ",t (s)∈Pk \Γ } `k ( ȳ",t (s), ᾱ",t (s))ds
is at a distance to FL+k (x) of the order of o" (1) + O(t). Therefore, we get that
Z
t
1{ ȳ",t (s)∈Pk \Γ } `k ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pk )( ȳ",t (s)). f k ( ȳ",t (s), ᾱ",t (s)) ds
τ",t
€
Š
≥ |{s : ȳ",t (s) ∈ Pk \Γ }| − max(ξ,ζ)∈FL+k (x) {−Dϕ(x, ξ) − ζ} + o" (1) + O(t) .
(3.5.26)
→ From Point 3 of Theorem 3.3.1, we know that f ( ȳ",t (s), ᾱ",t (s)) ∈ Re0 almost everywhere on
{s : ȳ",t (s) ∈ Γ }. Therefore,
‚Z t
Œ
Z t
1
1{ ȳ",t (s)∈Γ } f ( ȳ",t (s), ᾱ",t (s))ds,
1{ ȳ",t (s)∈Γ } `( ȳ",t (s), ᾱ",t (s))ds is at a dis|{s : ȳ",t (s) ∈ Γ }|
τ
τ
",t
",t
3.5. APPENDIX
119
¦
©
tance to co ∪Nj=1 FL j (x) ∩ (Re0 × R) of the order of o" (1) + O(t) and then
Z t
1{ ȳ",t (s)∈Γ } `( ȳ",t (s), ᾱ",t (s)) + DΓ ϕ( ȳ",t (s)). f ( ȳ",t (s), ᾱ",t (s)) ds
τ",t
≥ |{s : ȳ",t (s) ∈ Γ }| − max(ξ,ζ)∈co¦∪N FL (x)∩(Re ×R)© {−Dϕ(x, ξ) − ζ} + o" (1) + O(t) .
j=1
0
j
(3.5.27)
However, from the piecewise linearity of the function (ξ, ζ) 7→ −Dϕ(x, ξ) − ζ we have that
© {−Dϕ(x, ξ) − ζ}
¦ max
(ξ,ζ)∈co ∪Nj=1 FL j (x)∩(Re0 ×R)
= max
j∈{1,...,N }
max
(ξ,ζ)∈FL j (x)∩(Re0 ×R)
{−Dϕ(x, ξ) − ζ} .
Therefore, (3.5.27) give us finally
Z t
1{ ȳ",t (s)∈Γ } `( ȳ",t (s), ᾱ",t (s)) + DΓ ϕ( ȳ",t (s)). f ( ȳ",t (s), ᾱ",t (s)) ds
τ",t
≥ |{s : ȳ",t (s) ∈ Γ }| − max
j∈{1,...,N }
max
(ξ,ζ)∈FL j (x)∩(Re0 ×R)
(3.5.28)
τ",t
Z
→ It remains to deal with
0
{−Dϕ(x, ξ) − ζ} + o" (1) + O(t) .
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )( ȳ",t (s)). f i ( ȳ",t (s), ᾱ",t (s))ds. It is the
Z τ",t
f i ( ȳ",t (s), ᾱ",t (s)).ei ds = −x " .ei < 0 generates some out-
term the most tricky one because
0
going directions. To conclude, we have to use that x " .ei = o" (1). As a consequence, up to the
|x " .ei |
|x " .ei |
extraction of a subsequence, we may assume that either lim
= 0 or C1 ≤
≤ C2 , for
"→0 τ",t
τ",t
all ", for some positive constants C1 , C2 .
Z τ",t
Z τ",t
|x " .ei |
1
If lim
= 0, it is simple to see that
`i ( ȳ",t (s), ᾱ",t (s))ds
f i ( ȳ",t (s), ᾱ",t (s))ds,
"→0 τ",t
τ",t
0
0
is at a distance to FLi (x) ∩ (Re0 × R) of the order of o" (1) + O(t). Then, we get
Z τ",t
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )( ȳ",t (s)). f i ( ȳ",t (s), ᾱ",t (s))ds
0
≥ τ",t
If C1 ≤
Z
(3.5.29)
− max(ξ,ζ)∈FLi (x)∩(Re0 ×R) {−Dϕ(x, ξ) − ζ} + o" (1) + O(t) .
|x " .ei |
≤ C2 , we can directly prove that
τ",t
τ",t
`i ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|Pi )( ȳ",t (s)). f i ( ȳ",t (s), ᾱ",t (s))ds = O(x " .ei ) = o" (1).
(3.5.30)
0
Finally, if we put together all the informations of (3.5.25), (3.5.26), (3.5.28), (3.5.29), (3.5.30)
and the fact that ϕ(x " ) − ϕ(x) = o" (1), the inequality (3.5.22) gives us
0 ≥ −t max {
max
{−D(ϕ|P j )(x).ξ − ζ}} + v? (x)(e−λt − 1) + t o" (1) + O(t 2 ) + o" (1).
j∈{1,...,N } (ξ,ζ)∈FL+j (x)
Dividing this last inequality by t, taking the limit as " tends to 0 and in a second time the limit as
t tends to 0 we get the wanted inequality
λv? (x) +
max {−Dϕ(x, ξ) − ζ} ≥ 0.
(ξ,ζ)∈FL(x)
• If x ∈ Γ and x " ∈ Γ : In this case, the inequality (3.5.22) can be written as follow
N Z t
X
”
—
0 ≥ +
1{ ȳ",t (s)∈P j \Γ } ` j ( ȳ",t (s), ᾱ",t (s)) + D(ϕ|P j )( ȳ",t (s)). f j ( ȳ",t (s), ᾱ",t (s)) ds
+
Zj=1t
0
0
1{ ȳ",t (s)∈Γ } `( ȳ",t (s), ᾱ",t (s)) + DΓ ϕ( ȳ",t (s)). f ( ȳ",t (s), ᾱ",t (s)) ds
+v? (x)(e−λt − 1)+ϕ(x " ) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1).
120
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
The same study term by term that in the previous case gives us the desired result.
Show that v ? is a subsolution of (3.2.15) : for x ∈ S let (x " )">0 be a sequence such that x " tends
to x as " tends to 0 and v(x " ) tends to v ? (x) as " tends to 0. Let ϕ be in R(S ) such that v ? − ϕ has a
local maximum at x, i.e. there exists r > 0 such that
v ? (x) − ϕ(x) ≥ v ? ( y) − ϕ( y).
∀ y ∈ B(x, r) ∩ S ,
(3.5.31)
To prove that v ? is a subsolution, according to Corollary 3.3.1, it is enough to show that for any k ∈
{1, . . . , N },
v ? (x) +
sup
(−D(ϕ|Pk )(x). f k (x, a) − `k (x, a)) ≤ 0.
(3.5.32)
a∈Ak s.t. f k (x,a).ek >0
Let ā ∈ Ak be such that f k (x, a).ek = δā > 0.
• If x " ∈ Pi \ Γ : For all " > 0, let ( ȳ" , ᾱ" ) ∈ T x " be an admissible controlled trajectory given
by Lemma 3.3.2 and τ̄" (≤ C x " .ei ) its exit time from Pi \ Γ . So, we consider ỹ : [0, t̃) → Pk
Rt
the maximal solution of the integral equation ȳ(t) = ȳ" (τ̄" ) + 0 f k ( ȳ(s), ā)ds. According to
e 3], we can check that t̃ ≥ δā . Then, we introduce ( y" , α" ) the
the assumptions [H0] and [H
2L f M f
admissible controlled trajectory of T x " defined in [0, τ̄" + t̃) as follow
( y" (t), α" (t)) =
§
( ȳ" (t), ᾱ" (t))
( ȳ(t − τ̄" ), ᾱ)
if
if
t < τ̄" ,
T ≤ t ≥ τ̄" .
Note that by construction, for t ∈ (τ̄" , t̃) small enough, we have y" (s) ∈ Pi \ Γ for s ∈ (0, τ̄" ) and
y" (s) ∈ Pk \ Γ for s ∈ (τ̄" , t). In the sequel we will consider such a t.
From the dynamic programming principle, Proposition 3.2.1, we have
Rt
v(x " ) ≤
`( y" (s), α" (s))e−λs ds + e−λt v( y" (t))
R0t
≤
`( y" (s), α" (s))e−λs ds + e−λt v ? ( y" (t)).
0
Then, according to (3.5.31), for " small enough we have
Z t
`( y" (s), α" (s))e−λs ds + v ? (x)(e−λt − 1) + (ϕ( y" (t)) − ϕ(x))e−λt .
v(x " ) − v ? (x) ≤
0
By similar arguments as above, we can deduce from this inequality the following
Z t
`( y",t (s), α",t (s))ds + v ? (x)(e−λt − 1) + ϕ( y",t (t)) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1).
0≤
0
And by construction of the admissible controlled trajectory ( y" , α" ), this inequality can be written
as follows
Z τ̄"
`i ( y" (s), α" (s)) + D(ϕ|Pi )( y" (s)). f i ( y" (s), α" (s))ds
≤
0
0Z
t
+
τ̄"
`k ( y" (s), ā) + D(ϕ|Pk )( y" (s)). f k ( y" (s), ā)ds
(3.5.33)
+v ? (x)(e−λt − 1)+ϕ(x " ) − ϕ(x) + t o" (1) + O(t 2 ) + o" (1).
Z
τ̄"
So, using that
`i ( y" (s), α" (s)) + D(ϕ|Pi )( y" (s)). f i ( y" (s), α" (s))ds = o" (1)
0
1
and
t − τ̄"
Z
t
τ̄"
`k ( y" (s), ā) + D(ϕ|Pk )( y" (s)). f k ( y" (s), ā)ds
= D(ϕ|Pk )(x). f k (x, ā) + `k (x, ā) + o" (1) + O(t). Thus, the inequality (3.5.33) gives us
0 ≥ v ? (x)(e−λt − 1) + t D(ϕ|Pk )(x). f k (x, ā) + `k (x, ā) + t ◦" (1) + O(t 2 ) + o" (1).
3.5. APPENDIX
121
Dividing this last inequality by t, taking the limit as " tends to 0 and in a second time the limit as
t tends to 0 we give us
λv ? (x) − D(ϕ|Pk )(x). f k (x, ā) − `k (x, ā) ≤ 0.
Since this inequality is true for any ā ∈ Ak such that f k (x, a).ek > 0, we finally get (3.5.32).
• If x " ∈ Γ : Then, the same argument as above can be used. The only difference is that for
the construction of the admissible controlled trajectory ( y" , α" ) we do not need to join x " to Γ .
Consequently, the calculations that follow are slightly simpler.
122
CHAPITRE 3. THE MULTIDIMENSIONAL CASE
TROISIÈME
PARTIE
A singular perturbation problem
123
124
CHAPITRE 4
ASYMPTOTIC BEHAVIOUR OF
HAMILTON-JACOBI EQUATIONS DEFINED
ON TWO DOMAINS SEPARATED BY AN
OSCILLATING INTERFACE
Résumé
On considère une famille de problèmes de contrôle optimal dans le plan pour
lesquels les dynamiques et les coûts instantanés peuvent être discontinus à
travers une frontière oscillante Γ" . La frontière oscillante Γ" , de période et
d’amplitude de l’ordre de grandeur de ", tend vers une frontière droite Γ
quand " tend vers 0. On prouve que la fonction valeur v" tend vers l’unique
solution d’une équation de Hamilton-Jacobi effective comprenant une condition de transmission effective sur la frontière droite Γ . Cette condition de
transmission effective est la trace des oscillations des frontières Γ" .
Abstract
We consider a family of optimal control problems in the plane with dynamics
and running costs possibly discontinuous across an oscillatory interface Γ" .
The oscillations of the interface have small period and amplitude, both of
the order of ", and the interfaces Γ" tend to a straight line Γ . We study the
asymptotic behavior as " → 0. We prove that the value function tends to the
solution of Hamilton-Jacobi equations in the two half-planes limited by Γ ,
with an effective transmission condition on Γ keeping track of the oscillations
of Γ" .
4.1
Introduction
The goal of this paper is to study the asymptotic behavior as " → 0 of the value function of an optimal
control problem in R2 in which the running cost and dynamics may jump across a periodic oscillatory
interface Γ" , when the oscillations of Γ" have a small amplitude and period, both of the order of ". The interface Γ" separates two unbounded regions of R2 , Ω"L and ΩR" . To characterize the optimal control problem, one has to specify the admissible dynamics at a point x ∈ Γ" : in our setting, no mixture is allowed at
the interface, i.e. the admissible dynamics are the ones corresponding to the subdomain Ω"L and entering Ω"L , or corresponding to the subdomain ΩR" and entering ΩR" . Hence the situation differs from those
studied in the articles of G. Barles, A. Briani and E. Chasseigne [BBC13, BBC14] and of G. Barles, A.
Briani, E. Chasseigne and N. Tchou [BBCT14], in which mixing is allowed at the interface. The optimal
control problem under consideration has been first studied in [Oud14]: the value function is characterized as the viscosity solution of a Hamilton-Jacobi equation with special transmission conditions on Γ" ; a
comparison principle for this problem is proved in [Oud14] with arguments from the theory of optimal
control similar to those introduced in [BBC13, BBC14]. In parallel to [Oud14], Imbert and Monneau
have studied similar problems from the viewpoint of PDEs, see [IM14b], and have obtained comparison
results for quasi-convex Hamiltonians. In particular, [IM14b] contains a characterization of the viscosity solution of the transmission problem with a reduced set of test-functions; this characterization will
be used in the present work. Note that [Oud14,IM14b] can be seen as extensions of articles devoted to
the analysis of Hamilton-Jacobi equations on networks, see [ACCT13,IMZ13,AOT15b,IM14a], because
125
126
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
the notion of interface used there can be seen as a generalization of the notion of vertex (or junction)
for a network.
We will see that as " tends to 0, the value function converges to the solution of an effective problem
related to a flat interface Γ , with Hamilton-Jacobi equations in the half-planes limited by Γ and a transmission condition on Γ .
Whereas the partial differential equation far from the interface is unchanged, the main difficulty consists in finding the effective transmission condition on Γ . Naturally, the latter depends on the dynamics
and running cost but also keeps memory of the vanishing oscillations. The present work is closely
related to two recent articles, [AT15] and [GIM14], about singularly perturbed problems leading to
effective Hamilton-Jacobi equations on networks. Indeed, an effective Hamiltonian corresponding to
trajectories staying close to the junction was first obtained in [AT15] as the limit of a sequence of ergodic constants corresponding to larger and larger bounded subdomains. This construction was then
used in [GIM14] in a different case. Let us briefly describe the singular perturbation problems studied in [AT15] and [GIM14]: in [AT15], some of the authors of the present paper study a family of
star-shaped planar domains D" made of N non intersecting semi-infinite strips of thickness " and of a
central region whose diameter is proportional to ". As " → 0, the domains D" tend to a network G made
of N half-lines sharing an endpoint O, named the vertex or junction point. For infinite horizon optimal
control problems in which the state is constrained to remain in the closure of D" , the value function
tends to the solution of a Hamilton-Jacobi equation on G , with an effective transmission condition at
O. In [GIM14], Galise, Imbert and Monneau study a family of Hamilton-Jacobi equations in a simple
network composed of two half-lines with a perturbation of the Hamiltonian localized in a small region
close to the junction.
In the proof of convergence, we will see that the main technical point lies in the construction of correctors and in their use in the perturbed test-function method of Evans, see [Eva89]. As in [AT15]
and [GIM14], an important difficulty comes from the unboundedness of the domain in which the correctors are defined. The strategies for passing to the limit in [AT15] and [GIM14] differ: the method
proposed in [AT15] consists of contructing an infinite family of correctors related to the vertex, while
in [GIM14], only one corrector related to the vertex is needed thanks to the use of the above mentioned reduced set of test-functions. Arguably, the strategy proposed in [AT15] is more natural and
that in [GIM14] is simpler. For this reason, the technique implemented in the present work for proving
the convergence to the effective problem will be closer to the one proposed in [GIM14]. Note that
similar techniques are used in the very recent work [FS14a], which deals with applications to traffic
flows. The question of the correctors in unbounded domains has recently been addressed by P-L. Lions
in his lectures at Collège de France, [Lio14], precisely in january and february 2014: the lectures dealt
with recent and still unpublished results obtained in collaboration with T. Souganidis on the asymptotic
behavior of solutions of Hamilton-Jacobi equations in a periodic setting with some localized defects.
Finally, we stress the fact that the technique proposed in the present work is not specific to the transmission condition imposed on Γ" .
The paper is organized as follows: in the remaining part of § 4.1, we set the problem and give the
main result. In Section 4.2, we show that the problem is equivalent to a more convenient one, set
in a straightened fixed geometry. In § 4.3, we study the asymptotic behavior far from the interface
and introduce some ingredients that will be useful to define the effective transmission condition. In
§ 4.4, we define the effective cost/Hamiltonian for moving along the effective interface, and related
correctors. This is of course a key step in the study of the asymptotic behavior. Section 4.5 deals with
further properties of the correctors, in particular their growth at infinity. The comparison result for the
effective problem is stated in § 4.6, and the proof of the main convergence theorem is written in § 4.7.
4.1. INTRODUCTION
4.1.1
127
The geometry
 ‹
 ‹
1
0
, e2 =
. Let g : R → R be a C 2 -function,
0
1
periodic with period 1. For any " > 0, let (Ω"L , Γ" , ΩR" ) be the partition of R2 defined as follows:
n
x o
2
,
(4.1.1)
Γ" =
(x 1 , x 2 ) ∈ R2 : x 1 = " g
"
n
x o
2
Ω"L =
(x 1 , x 2 ) ∈ R2 : x 1 < " g
,
(4.1.2)
"
n
x o
2
ΩR" =
(x 1 , x 2 ) ∈ R2 : x 1 > " g
.
(4.1.3)
"
Let (e1 , e2 ) be an orthonormal basis of R2 : e1 =
Note that ∂ Ω"L = ∂ ΩR" = Γ" . For x ∈ Γ" , the vector
n" (x) =
‹
1
x
−g 0 ( "2 )

(4.1.4)
is normal to Γ" and oriented from Ω"L to ΩR" . Defining also
§
1
if i = R,
σi =
−1 if i = L,
(4.1.5)
the vector σ i n" (x) is normal to Γ" at the point x ∈ Γ" and points toward Ω"i .
The geometry obtained at the limit when " → 0 can also be found by taking g = 0 in the definitions
above: let (Ω L , Γ , ΩR ) be the partition of R2 defined by
Γ =
(x 1 , x 2 ) ∈ R2 : x 1 = 0 ,
(4.1.6)
L
2
Ω
=
(x 1 , x 2 ) ∈ R : x 1 < 0) ,
(4.1.7)
R
2
Ω
=
(x 1 , x 2 ) ∈ R : x 1 > 0) .
(4.1.8)
One sees that ∂ Ω L = ∂ ΩR = Γ and that for all x ∈ Γ , the unit normal vector to Γ at x pointing toward
ΩR is n(x) = e1 . The two kinds of geometry are represented in Figure 4.1.
Γ"
Ω"L
ΩR"
x•
Γ
Ω
ΩR
L
n" (x)
(a)
(b)
Figure 4.1: (a): Γ" is an oscillating interface with an amplitude and period of " . (b): the geometry
obtained at the limit when " → 0
4.1.2
The optimal control problem in Ω"L ∪ ΩR" ∪ Γ"
We consider infinite-horizon optimal control problems which have different dynamics and running costs
in the regions Ω"i , i = L, R. The sets of controls associated to the index i = L, R will be called Ai ; similarly,
the notations f i and `i will be used for the dynamics and running costs. The following assumptions
will be made in all the work
4.1.2.1
Standing Assumptions
[H0] A is a metric space (one can take A = Rm ). For i = L, R, Ai is a non empty compact subset of A
and f i : R2 × Ai → R2 is a continuous bounded function. The sets Ai are disjoint. Moreover, there
128
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
exists L f > 0 such that for any i = L, R, x, y ∈ R2 and a ∈ Ai ,
| f i (x, a) − f i ( y, a)| ≤ L f |x − y|.
Define M f = maxi=L,R sup x∈R2 ,a∈Ai | f i (x, a)|. The notation F i (x) will be used for the set F i (x) =
{ f i (x, a), a ∈ Ai }.
[H1] For i = L, R, the function `i : R2 × Ai → R is continuous and bounded. There is a modulus of
continuity ω` such that for any i = L, R, x, y ∈ R2 and a ∈ Ai ,
|`i (x, a) − `i ( y, a)| ≤ ω` (|x − y|).
Define M` = maxi=L,R sup x∈R2 ,a∈Ai |`i (x, a)|.
[H2] For any i = L, R and x ∈ R2 , the non empty set FLi (x) = {( f i (x, a), `i (x, a)), a ∈ Ai } is closed
and convex.
[H3] There is a real number δ0 > 0 such that for i = L, R and all x ∈ Γ" , B(0, δ0 ) ⊂ F i (x).
We stress the fact that all the results below hold provided the latter assumptions are satisfied, although,
in order to avoid tedious repetitions, we will not mention them explicitly in the statements.
We refer to [AOT15b] and [Oud14] for comments on the assumptions and the genericity of the model.
4.1.2.2
The optimal control problem
Let the closed set M" be defined as follows:
M" = (x, a); x ∈ R2 , a ∈ Ai if x ∈ Ω"i , i = L, R, and a ∈ AL ∪ AR if x ∈ Γ" .
(4.1.9)
The dynamics f" is a function defined in M" with values in R2 :
∀(x, a) ∈ M" ,
f" (x, a) =
§
f i (x, a)
f i (x, a)
if x ∈ Ω"i ,
if x ∈ Γ" and a ∈ Ai .
The function f" is continuous on M" because the sets Ai are disjoint. Similarly, let the running cost
`" : M" → R be given by
∀(x, a) ∈ M" ,
`" (x, a) =
§
`i (x, a)
`i (x, a)
if x ∈ Ω"i ,
if x ∈ Γ" and a ∈ Ai .
For x ∈ R2 , the set of admissible trajectories starting from x is

∞
+
y x ∈ Lip(R+ ; R2 ),
 ( y x , a) ∈ Lloc (R ; M" ) :
Z t
T x," =
y x (t) = x +
f" ( y x (s), a(s))ds

0


∀t ∈ R+ 
.
(4.1.10)
The cost associated to the trajectory ( y x , a) ∈ T x," is
J" (x; ( y x , a)) =
Z
∞
`" ( y x (t), a(t))e−λt d t,
(4.1.11)
0
with λ > 0. The value function of the infinite horizon optimal control problem is
v" (x) =
inf
( y x ,a)∈T x,"
J" (x; ( y x , a)).
Proposition 4.1.1. The value function v" is bounded and continuous in R2 .
Proof. This result is classical and can be proved with the same arguments as in [BCD97].
(4.1.12)
4.1. INTRODUCTION
4.1.3
129
The Hamilton-Jacobi equation
Similar optimal control problems have recently been studied in [AOT15b, IM14a, Oud14, IM14b]. It
turns out that v" can be characterized as the viscosity solution of a Hamilton-Jacobi equation with a
discontinuous Hamiltonian, (once the notion of viscosity solution has been specially tailored to cope
with the above mentioned discontinuity). We briefly recall the definitions used e.g. in [Oud14].
4.1.3.1
Test-functions
Definition 4.1.1. For " > 0, the function φ : R2 → R is an admissible (")-test-function if φ is continuous
in R2 and for any i ∈ {L, R}, φ|Ωi ∈ C 1 (Ω"i ).
"
The set of admissible test-functions is noted R" . If φ ∈ R" , x ∈ Γ" and i ∈ {L, R}, we set Dφ i (x) =
lim x 0 →x Dφ(x 0 ).
x 0 ∈Ω"i
4.1.3.2
Hamiltonians
For i = L, R, let the Hamiltonians H i : R2 × R2 → R and HΓ" : Γ" × R2 × R2 → R be defined by
H i (x, p)
=
max(−p · f i (x, a) − `i (x, a)),
HΓ" (x, p L , pR )
=
max{ HΓ" (x, p L ), HΓ" (x, pR )},
a∈Ai
+,R
+,L
(4.1.13)
(4.1.14)
where, with n" (x) and σ i defined in § 4.1.1,
+,i
HΓ" (x, p) =
4.1.3.3
a∈Ai
s.t.
max
σ i f i (x,a)·n" (x)≥0
(−p · f i (x, a) − `i (x, a)),
∀x ∈ Γ" , ∀p ∈ R2 .
(4.1.15)
Definition of viscosity solutions
We now recall the definition of a viscosity solution of
λu + H" (x, Du) = 0.
(4.1.16)
Definition 4.1.2.
• An upper semi-continuous function u : R2 → R is a subsolution of (4.1.16) if for
2
any x ∈ R , any φ ∈ R" s.t. u − φ has a local maximum point at x, then
λu(x) + H i (x, Dφ i (x))
≤ 0,
if x ∈ Ω"i ,
(4.1.17)
λu(x) + HΓ" (x, Dφ (x), Dφ (x))
≤ 0,
if x ∈ Γ" ,
(4.1.18)
L
R
see Definition § 4.1.1 for the meaning of Dφ i (x) if x ∈ Γ" .
• A lower semi-continuous function u : R2 → R is a supersolution of (4.1.16) if for any x ∈ R2 , any
φ ∈ R" s.t. u − φ has a local minimum point at x, then
λu(x) + H i (x, Dφ i (x))
≥ 0,
if x ∈ Ω"i ,
(4.1.19)
λu(x) + HΓ" (x, Dφ (x), Dφ (x))
≥0
if x ∈ Γ" .
(4.1.20)
L
R
• A continuous function u : R2 → R is a viscosity solution of (4.1.16) if it is both a viscosity sub and
supersolution of (4.1.16).
4.1.3.4
Characterization of v" as a viscosity solution of (4.1.16)
The following theorem will be proved below, see Theorem 4.2.3, by finding an equivalent optimal
control problem in a straightened fixed geometry and using results proved in [Oud14]:
Theorem 4.1.1. The value function v" defined in (4.1.12) is the unique bounded viscosity solution of
(4.1.16).
130
4.1.4
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Main result and organization of the paper
We now state our main result:
Theorem 4.1.2. As " → 0, v" converges uniformly to v the unique bounded viscosity solution of
λv(z) + H i (z, Dv(z)) = 0
λv(z) + max E(z2 , ∂z2 v(z)), HΓ (z, Dv L (z), Dv R (z)) = 0
if z ∈ Ωi ,
(4.1.21)
if z = (0, z2 ) ∈ Γ ,
(4.1.22)
which we note for short
λv + H (x, Dv) = 0.
(4.1.23)
i
The Hamiltonians H , HΓ and E are respectively defined in (4.1.13), (4.3.7) below, and (4.4.18) below.
Let us list the notions which are needed by Theorem 4.1.2 and give a few comments:
1. Problem (4.1.23) is a transmission problem across the interface Γ , with the effective transmission
condition (4.1.22). The notion of viscosity solutions of (4.1.23) is similar to the one proposed in
Definition 4.1.2, replacing Γ" with Γ .
2. Note that the Hamilton-Jacobi equations in Ω L and ΩR are directly inherited from (4.1.17): this is
quite natural, since the interface Γ" oscillates with an amplitude of the order of ", which therefore
vanishes as " → 0.
3. The Hamiltonian HΓ appearing in the effective transmission condition at the junction is defined
in § 4.3.2, precisely in (4.3.7); it is built by considering only the dynamics related to Ωi which
point from Γ toward Ωi , for i = L, R.
4. The effective Hamiltonian E is the only ingredient in the effective problem that keeps track of the
oscillations of Γ" , i.e. of the function g. It is constructed in § 4.4, see (4.4.18), as the limit of
a sequence of ergodic constants related to larger and larger bounded subdomains. This is reminiscent of a construction first performed in [AT15] for singularly perturbed problems in optimal
control leading to Hamilton-Jacobi equations posed on a network. A similar construction can also
be found in [GIM14].
5. The references [AT15] and [GIM14] propose two different possible strategies for proving Theorem 4.1.2: in this work, the chosen strategy is reminiscent of [GIM14], because it relies on the
construction of a single corrector, whereas the method proposed in [AT15] requires the construction of an infinite family of correctors. This will be done in § 4.4 and the slopes at infinity of the
correctors will be studied in § 4.5.
4.2
Straightening the geometry
It will be convenient to use a change of variables depending on " and set the problem in a straightened
and fixed geometry.
4.2.1
A change of variables
The following change of variables
can be used
in ‹a fixed geometry:
‹ to write the optimal control
 problem

x
x
x 1 + " g( "2 )
x 1 − " g( "2 )
−1
2
. We see that G (x) =
. The oscillatory
for x ∈ R , take z = G(x) =
x2
x2
interface Γ" is mapped onto Γ = {z : z1 = 0} by G. The Jacobian of G is

x ‹
1 −g 0 ( "2 )
J" (x) =
,
(4.2.1)
0
1
x
1 g 0 "2
−1
and it inverse is J" (x) =
. The following properties will be useful: for any x ∈ R2 ,
0
1
sup
X ∈R2 ,|X |≤1
J" (G −1 (x)) = J" (x)
p
|J" (x)X | ≤ 2(1+ k g 0 k∞ )
and
and
J"−1 (G −1 (x)) = J"−1 (x),
(4.2.2)
p
−1
0
sup |J" (x)X | ≤ 2(1+ k g k∞ ), (4.2.3)
X ∈R2 ,|X |≤1
4.2. STRAIGHTENING THE GEOMETRY
131
where |·| stands for the euclidean norm. Note that (4.2.2) holds because G and G −1 leave x 2 unchanged,
and J" only depends on x 2 .
4.2.2
The optimal control problem in the straightened geometry
For i = L, R, we define the new dynamics f˜"i and running costs ˜`"i as
f˜"i :
Ω̄i × Ai
(z, a)
→
7
→
R2
J" (z) f i G −1 (z), a ,
(4.2.4)
˜`i :
"
Ω̄i × Ai
(z, a)
→
7→
R
`i G −1 (z), a .
(4.2.5)
We deduce the following properties from the standing assumptions [H0]-[H3]:
e 0]" For i = L, R and " > 0, the function f˜i is continuous and bounded. Moreover, there exists
[H
"
L̃ f (") > 0 and M̃ f > 0 such that for any z, z 0 ∈ Ω̄i and a ∈ Ai ,
| f˜"i (z, a) − f˜"i (z 0 , a)| ≤ L̃ f (")|z − z 0 |
and
| f˜"i (z, a)| ≤ M̃ f .
e 1]" For i = L, R and " > 0, the function ˜`i is continuous and bounded. Moreover, if we set ω̃` (t) =
[H
"
p
ω` ( 2(1+ k g 0 k∞ )t), then for any z, z 0 ∈ Ω̄i and a ∈ Ai ,
|˜`"i (z, a) − ˜`"i (z 0 , a)| ≤ ω̃` (|z − z 0 |),
and
|˜`"i (x, a)| ≤ M` ,
the constants M` and the modulus of continuity ω` (·) being introduced in [H1].
˜ i (x) = {( f˜i (x, a), ˜`i (x, a)), a ∈ Ai } is
e 2]" For any i = L, R, " > 0 and x ∈ Ω̄i , the non empty set FL
[H
"
"
"
closed and convex.
e 3]" For any i = L, R and " > 0, if we set δ̃0 =
[H
{ f˜i (z, a), a ∈ Ai }.
δ0
p
,
2(1+kg 0 k∞ )
then for any z ∈ Γ , B(0, δ̃0 ) ⊂ F̃"i (z) =
"
e 0]" and [H
e 1]" result from direct calculations. Property [H
e 2]" comes from the fact that
Properties [H
e 3]" , take i = L, R, z = (0, z2 ) ∈ Γ and
linear maps preserve the convexity property. In order to prove [H
p
p ∈ B(0, δ̃0 ). We look for a ∈ Ai such that f˜"i (z, a) = p. Using (4.2.3), we see that |J"−1 (z)p| ≤ 2(1+ k
g 0 k∞ )δ̃0 ≤ δ0 . Thus from [H3], since G −1 (z) ∈ Γ" , there exists ā ∈ Ai such that f i G −1 (z), ā =
J"−1 (z)p, and we obtain that f˜"i (z, ā) = p.
Let us now define the counterparts of M" , f" and `" :
M = (x, a); x ∈ R2 , a ∈ Ai if x ∈ Ωi , and a ∈ AL ∪ AR if x ∈ Γ ,
(4.2.6)
i
i
˜
f" (z, a)
if x ∈ Ω ,
∀(z, a) ∈ M ,
f˜" (z, a) =
(4.2.7)
f˜"i (z, a)
if x ∈ Γ and a ∈ Ai ,
˜`i (z, a)
if x ∈ Ωi ,
˜
∀(z, a) ∈ M ,
`" (z, a) = ˜"i
(4.2.8)
`" (z, a)
if x ∈ Γ and a ∈ Ai .
For x ∈ R2 , the set of admissible trajectories starting from x is

∞
+
y x ∈ Lip(R+ ; R2 ),
 ( y x , a) ∈ Lloc (R ; M ) :
Z t
Tex," =
y x (t) = x +
f˜" ( y x (s), a(s))ds

0


∀t ∈ R+ 
.
(4.2.9)
132
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Note that ∀z ∈ R2 , ( yz , a) ∈ Tez," ⇔ G −1 ( yz (·)), a ∈ TG −1 (z)," . The new optimal control problem
consists in finding
Z∞
˜`" ( yz (t), a(t))e−λt d t.
ṽ" (z) = inf
(4.2.10)
( yz ,a)∈Tez,"
0
Remark 4.2.1 (Relationship between v" and ṽ" ). For any z ∈ R2 ,
z ṽ" (z) = v" (G −1 (z)) = v" z1 + " g 2 , z2 .
"
4.2.3
The Hamilton-Jacobi equation in the straightened geometry
4.2.3.1
Hamiltonians
If i ∈ {L, R}, the Hamiltonians H̃"i : R2 × R2 → R are defined by
H̃"i (z, p) = max − f˜"i (z, a) · p − ˜`"i (z, a) = max −J" (z) f i (G −1 (z), a) · p − `i (G −1 (z), a) .
a∈Ai
a∈Ai
(4.2.11)
More explicitly,
 
1
H̃"i (z, p) = max −
0
a∈Ai
‹
z ‹
z
z
−g 0 ( "2 ) i
f ((z1 + " g( 2 ), z2 ), a) · p − `i ((z1 + " g( 2 ), z2 ), a) .
1
"
"
If z ∈ Γ , the Hamiltonian H̃Γ ," : Γ × R2 × R2 → R is defined by
+,L
+,R
H̃Γ ," (z, p L , pR ) = max H̃Γ ," (z, p L ), H̃Γ ," (z, pR ) ,
(4.2.12)
where for i = 1, 2, z ∈ R2 , p i ∈ R2 , and σ i is defined in § 4.1.1,
+,i
H̃Γ ," (z, p) =
max
(− f˜"i (z, a) · p − ˜`"i (z, a)).
i
a ∈ A s.t.
σ i f˜"i (z, a) · e1 ≥ 0
+,i
+,i
Note that if z ∈ Γ , then n" (G −1 (z)) = J" (z) T e1 , by (4.2.2). Hence, H̃Γ ," is the counterpart of HΓ ," .
4.2.3.2
Definition of viscosity solutions in the straightened geometry
Definition 4.2.1. The function φ : R2 → R is an admissible test-function for the fixed geometry if φ is
continuous in R2 and for any i = L, R, φ|Ω̄i ∈ C 1 (Ω̄i ).
The set of the admissible test-functions is denoted R. If φ ∈ R, x ∈ Γ and i ∈ {L, R}, we set Dφ i (x) =
lim x 0 →x Dφ(x 0 ). Of course, the partial derivatives of φ|Ω̄ L and φ|Ω̄R with respect to x 2 coincide on Γ .
x 0 ∈Ωi
We then define the sub/super-solutions and solutions of
λu + H˜" (z, Du) = 0
(4.2.13)
as in Definition 4.1.2, using the set of test-functions R, the Hamiltonians H̃"i (z, p) if z ∈ Ωi and
H̃Γ ," (z, p L , pR ) if z ∈ Γ .
Remark 4.2.2. Let u : R2 → R be an upper semi-continuous (resp. lower semi-continuous) function and
ũ : R2 → R be defined by ũ(z) = u(G −1 (z)). Then u is a subsolution (resp. supersolution) of (4.1.16) if
and only if ũ is a subsolution (resp. supersolution) of (4.2.13).
4.2.3.3
Existence and uniqueness
We have seen in Remark 4.2.1 that the optimal control problems (4.1.12) and (4.2.10) are equivalent; similarly Remark 4.2.2 tells us that the notions of viscosity solutions of (4.1.16) and (4.2.13) are
equivalent. Therefore, it is enough to focus on (4.2.10) and (4.2.13).
4.3. ASYMPTOTIC BEHAVIOR IN Ω L AND ΩR
133
Lemma 4.2.1. There exists r > 0 such that any bounded viscosity subsolution u of (4.1.16) (resp.
λkuk∞ +M`
(resp.
(4.2.13)) is Lipschitz continuous in B(Γ" , r) (resp. B(Γ , r)) with Lipschitz constant Lu ≤
δ0
p (λkuk∞ +M` )(1+kg 0 k∞ )
2
Lu ≤ 2
), where for X a closed subset of R , B(X , r) denotes the set { y ∈ R2 :
δ0
dist( y, X ) < r}.
Proof. For a subsolution u of (4.2.13), the result is exactly [Oud14, Lemma 2.6].
If u is a subsolution of (4.1.16), then ũ(z) = u(G −1 (z)) is a subsolution of (4.2.13), and is therefore
Lipschitz continuous in a neighborhood of Γ . Since u = ũ◦G, u is Lipschitz continuous in a neighborhood
of Γ" .
Theorem 4.2.1 (Local comparison principle). Let u be a bounded viscosity subsolution of (4.1.16) (resp.
(4.2.13)), and v be a bounded viscosity supersolution of (4.1.16) (resp. (4.2.13)). For any z ∈ R2 , there
exists r > 0 such that
k (u − v)+ k L ∞ (B(z,r)) ≤k (u − v)+ k L ∞ (∂ B(z,r)) .
(4.2.14)
Proof. Let us focus on (4.2.13). If z ∈ Ωi , then we can choose r > 0 small enough so that B(z, r) ⊂ Ωi
and the result is classical. If z ∈ Γ , the result stems from a direct application of [Oud14, Theorem 3.3].
Indeed, all the assumptions required by [Oud14, Theorem 3.3] are satisfied thanks to the properties
e 0]" -[H
e 3]" . The result for (4.1.16) can be deduced from the latter thanks to Remark 4.2.2.
[H
Theorem 4.2.2 (Global comparison principle). Let u be a bounded viscosity subsolution of (4.1.16) (resp.
(4.2.13)), and v be a bounded viscosity supersolution of (4.1.16) (resp. (4.2.13)). Then u ≤ v.
Proof. The result for equation (4.2.13) stems from a direct application of [Oud14, Theorem 3.4]. Then
we deduce the result for equation (4.1.16) thanks to Remark 4.2.2.
Theorem 4.2.3. The value function v" (resp. ṽ" ) defined in (4.1.12) (resp. (4.2.10)) is the unique
bounded viscosity solution of (4.1.16) (resp. (4.2.13)).
Proof. Uniqueness is a direct consequence of Theorem 4.2.2 for both equations (4.1.16) and (4.2.13).
Existence for equation (4.2.13) is proved in the same way as in [Oud14, Theorem 2.3]. Then, existence
for (4.1.16) is deduced from Remark 4.2.2.
Remark 4.2.3. Under suitable assumptions, see §4 in [Oud14], all the above results hold if we modify
(4.1.16) (resp.(4.2.13)) by adding to HΓ" (resp. H̃Γ ," ) a Hamiltonian H"0 (resp. H̃"0 ) correponding to
trajectories staying on the junctions.
4.3
Asymptotic behavior in Ω L and ΩR
Our goal is to understand the asymptotic behavior of the sequence (v" )" as " tends to 0. In this section,
we are going to see that the Hamilton-Jacobi equations remain unchanged in Ω L and ΩR ; this is not
surprising because the amplitude of the oscillations of the interface vanishes as " → 0. Then, we are
going to introduce some of the ingredients of the effective boundary conditions on Γ .
From Remark 4.2.1, the sequence (v" )" converges if and only if the sequence (ṽ" )" converges. Moreover,
if they converge, the two sequences have the same limit. It will be convenient to focus on the asymptotic
behavior of the sequence (ṽ" )" , since the geometry is fixed. It is now classical to consider the relaxed
semi-limits
ṽ(z) = lim sup∗ ṽ" (z) = lim sup ṽ" (z 0 )
"
z 0 →z,"→0
and
ṽ(z) = lim inf∗ ṽ" (z) = lim
inf ṽ" (z 0 ).
0
"
Note that ṽ and ṽ are well defined, since (ṽ" )" is uniformly bounded by
z →z,"→0
M`
λ ,
see (4.2.10).
(4.3.1)
134
4.3.1
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
For the Hamilton-Jacobi equations in Ω L and ΩR , nothing changes
Proposition 4.3.1. For i = L, R, the functions ṽ(z) and ṽ(z) are respectively a bounded subsolution and
a bounded supersolution in Ωi of
λu(z) + H i (z, Du(z)) = 0,
(4.3.2)
where the Hamiltonian H i is given by (4.1.13).
Proof. The proof is classical but we give it for completeness. We only prove that ṽ(·) is a subsolution of
(4.3.2) in Ω L , the proof that ṽ(·) is a supersolution being similar. Take z̄ = (z̄1 , z̄2 ) ∈ Ω L , φ ∈ C 1 (Ω L )
and r0 > 0 such that B(z̄, r0 ) ⊂ Ω L and 0 = (ṽ − φ)(z̄) > (ṽ − φ)(z) for any z ∈ B(z̄, r0 ) \ {z̄}. We wish
to prove that λφ(z̄) + H L (z̄, Dφ(z̄)) ≤ 0. Suppose by contradiction that there exists η > 0 such that
λφ(z̄) + H L (z̄, Dφ(z̄)) = η.
(4.3.3)
z
Let us define φ" (z) = φ(z) + "∂z1 φ(z̄)g( "2 ) − δ.
We claim that for " > 0, δ > 0 and r > 0 small enough, φ" (z) is a supersolution of
łφ" (z) + H̃"L (z, Dφ" (z)) ≥
η
,
2
(4.3.4)
in B(z̄, r), where H̃"L is defined at (4.2.11). Indeed, φ" is a regular function which satisfies
z
H̃"L (z, Dφ" (z)) = H L G −1 (z), Dφ(z) + g 0 ( 2 ) ∂z1 φ(z̄) − ∂z1 φ(z) e2 .
"
Using the regularity of φ, ` L and f L and the boundedness of g and g 0 , this implies that
H̃"L (z, Dφ" (z))) = H L (z̄, Dφ(z̄)) + o|z−z̄|→0 (1) + o|"|→0 (1).
(4.3.5)
Then the claim (4.3.4) follows from (4.3.5) and (4.3.3).
The value of r can always be decreased so that 0 < r < r0 yielding that max∂ B(z̄,r) (ṽ − φ)(z) < 0. Fixing
r, for " small enough, max∂ B(z̄,r) (ṽ" − φ" )(z) ≤ 0. The local comparison principle in Theorem 4.2.1
implies that for any z ∈ B(z̄, r), (ṽ" − φ" )(z) ≤ 0. Letting " → 0, ṽ(z) ≤ φ(z) − δ for any z ∈ B(z̄, r).
This leads to ṽ(z̄) ≤ φ(z̄) − δ < ṽ(z̄), which cannot happen.
4.3.2
An ingredient in the effective transmission condition on Γ : the Hamiltonian HΓ inherited from the half-planes
For i ∈ {L, R}, let us define the Hamiltonian H +,i and H −,i : R2 × R2 → R by
H ±,i (z, p) =
a∈Ai
s.t.
max
(−p. f i (z, a) − `i (z, a)).
±σ i f i (z,a).e1 ≥0
(4.3.6)
and HΓ : Γ × R2 × R2 → R by
HΓ (z, p L , pR ) = max H +,L (z, p L ), H +,R (z, pR ) .
(4.3.7)
As in [AT15, GIM14], we introduce the functions E0i : R × R → R, i = L, R and E0 : R × R → R:
E0i (z2 , p2 )
=
E0 (z2 , p2 )
=
min H i ((0, z2 ), p2 e2 + qe1 ), q ∈ R ,
max E0L (z2 , p2 ), E0R (z2 , p2 ) .
(4.3.8)
(4.3.9)
The following lemma, which is the same as [Oud14, Lemma 2.1], deals with some monotonicity properties of H ±,i :
Lemma 4.3.1.
1. For any (0, z2 ) ∈ Γ , p ∈ R2 , p1 7→ H +,L ((0, z2 ), p + p1 e1 ) and p1 7→ H −,R ((0, z2 ), p + p1 e1 ) are
nondecreasing; p1 →
7 H −,L ((0, z2 ), p + p1 e1 ) and p1 7→ H +,R ((0, z2 ), p + p1 e1 ) are nonincreasing.
4.4. A NEW HAMILTONIAN INVOLVED IN THE EFFECTIVE TRANSMISSION CONDITION 135
−,L
+,L
2. For z2 , p2 ∈ R, there exist two unique real numbers p0 (z2 , p2 ) ≤ p0
H L ((0, z2 ), p2 e2 + pe1 )
H −,L ((0, z2 ), p2 e2 + pe1 ) =
E0L (z2 , p2 )
L
E0 (z2 , p2 )
H +,L ((0, z2 ), p2 e2 + pe1 ) =
H L ((0, z2 ), p2 e2 + pe1 )
+,R
4.4
−,L
if p ≤ p0 (z, p2 ),
−,L
if p > p0 (z, p2 ),
if p ≤ p0+L (z, p2 ),
+,L
if p > p0 (z, p2 ).
−,R
3. For z2 , p2 ∈ R, there exist two unique real numbers p0 (z2 , p2 ) ≤ p0
E0R (z2 , p2 )
−,R
H ((0, z2 ), p2 e2 + pe1 ) =
H R ((0, z2 ), p2 e2 + pe1 )
H R ((0, z2 ), p2 e2 + pe1 )
H +,R ((0, z2 ), p2 e2 + pe1 ) =
E0R (z2 , p2 )
(z2 , p2 ) such that
(z2 , p2 ) such that
−,R
if p ≤ p0 (z, p2 ),
−,R
if p > p0 (z, p2 ),
+,R
if p ≤ p0 (z, p2 ),
+,R
if p > p0 (z, p2 ).
A new Hamiltonian involved in the effective transmission condition
In this section, we construct the effective Hamiltonian corresponding to effective dynamics staying
on the interface Γ , by using similar ideas as those proposed in [AT15]. We will define an effective
Hamiltonian E on Γ as the limit of a sequence of ergodic constants for state-constrained problems
in larger and larger truncated domains. We will also construct correctors associated to the effective
Hamiltonian. The noteworthy difficulty is that the correctors need to be defined in an unbounded
domain.
4.4.1
Fast and slow variables
z
Let us introduce the fast variable y2 = "2 . Neglecting the contribution of " g( y2 ) in the Hamiltonians
H̃"i previously defined in (4.2.11), we obtain the new Hamiltonians H̃ i : R2 × R2 × R → R:
(4.4.1)
H̃ i (z, p, y2 ) := max −J̃( y2 ) f i (z, a) · p − `i (z, a) ,
a∈Ai
where

1
J̃( y2 ) =
0
‹
−g 0 ( y2 )
.
1
As above, using σ i introduced in § 4.1.1, we also define H̃ +,i and H̃ −,i : R2 × R2 × R → R by
H̃ ±,i (z, p, y2 ) =
max
−J̃( y2 ) f i (z, a) · p − `i (z, a) .
a ∈ Ai s.t.
i
±σ J̃( y2 ) f i (z, a) · e1 ≥ 0
(4.4.2)
(4.4.3)
Lemma 4.4.1. With L f , M f , M` and δ0 appearing in Assumptions [H0]-[H3], for any z ∈ R2 , y2 ∈ R
and p, p0 ∈ R2 ,
|H̃ i (z, p, y2 ) − H̃ i (z, p0 , y2 )| ≤ M f |p − p0 |, i = L, R,
(4.4.4)
and there exists a constant M > 0 (which can be computed from L f , M f , δ0 and kg 0 k∞ ) and a modulus of
continuity ω (which can be deduced from ω` , M` L f , δ0 and kg 0 k∞ ) such that for any z, z 0 ∈ R2 , y2 ∈ R
and p ∈ R2 ,
| H̃ i (z, p, y2 ) − H̃ i (z 0 , p, y2 ) |≤ M |p||z − z 0 | + ω(|z − z 0 |).
(4.4.5)
Similar estimates hold for H̃ +,i and H̃ −,i .
Proof. The proof is standard for the Hamiltonians H̃ i . Adapting the proofs of Lemmas 3.5 and 3.6
in [Oud14], we see that similar estimates hold for H̃ +,i and H̃ −,i : the proofs are not direct and rely
on the convexity of {(J̃( y2 ) f i (z, a), `i (z, a)), a ∈ Ai }, on the fact that the dynamics J̃( y2 ) f i (z, ·) satisfy
a strong controlability property similar to [H3] uniformly w.r.t. z, and on continuity properties of
(z, a) 7→ J̃( y2 ) f i (z, a) and (z, a) 7→ `i (z, a) similar to [H0] and [H1].
136
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Remark 4.4.1. Take z ∈ Ωi and p ∈ R2 . The unique real number λi (z, p) such that the following one
dimensional cell problem in the variable y2
§
H̃ i (z, p + χ 0 ( y2 )e2 , y2 ) = λi (z, p),
χ is 1-periodic w.r.t. y2 ,
(4.4.6)
admits a viscosity solution is λi (z, p) = H i (z, p); indeed, for this choice of λi (z, p), it is easy to check that
χ( y2 ) = p1 g( y2 ) is a solution of (4.4.6) and the uniqueness of λi (z, p) such that (4.4.6) has a solution is
well known, see e.g. [LPV, Eva89].
Remark 4.4.2. For any (0, z2 ) ∈ Γ , p ∈ R2 and y2 ∈ R, the functions p1 7→ H̃ ±,i ((0, z2 ), p + p1 e1 , y2 ) have
the same monotonicity properties as those stated in point 1 in Lemma 4.3.1 for p1 7→ H ±,i ((0, z2 ), p+ p1 e1 ).
Similarly, one can prove the counterparts of points 2 and 3 in Lemma 4.3.1 involving H̃ ±,i ((0, z2 ), p +
p1 e1 , y2 ), H̃ i ((0, z2 ), p + p1 e1 , y2 ) and
Ẽ0i (z2 , p2 , y2 ) = min H̃ i ((0, z2 ), p2 e2 + qe1 , y2 ), q ∈ R .
4.4.2
Ergodic constants for state-constrained problems in truncated domains
4.4.2.1
State-constrained problem in truncated domains
Let us fix z = (0, z2 ) ∈ Γ and p2 ∈ R. For ρ > 0, we consider the truncated cell problem:
 L
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 )
= λρ (z2 , p2 ) in (−ρ, 0) × R,



R

H̃
((0,
z
),
Du(
y)
+
p
e
,
y
)

2
2 2
2
= λρ (z2 , p2 ) in (0, ρ) × R,

 max H̃ +,i ((0, z ), Dui ( y) + p e , y ) = λ (z , p ) on Γ ,
2
2 2
2
ρ 2 2
i=L,R

H̃ −,L ((0, z2 ), Du( y) + p2 e2 , y2 )




H̃ −,R ((0, z2 ), Du( y) + p2 e2 , y2 )


u is 1-periodic w.r.t. y2 ,
= λρ (z2 , p2 )
= λρ (z2 , p2 )
on {−ρ} × R,
on {ρ} × R,
(4.4.7)
where the Hamiltonians H̃ i , H̃ +,i and H̃ −,i are respectively defined in (4.4.1) and (4.4.3). The notions of
viscosity subsolution, supersolution and solution of (4.4.7) are defined in the same way as in Definition
4.1.2 using the set of test-functions
Rρ = ψ|[−ρ,ρ]×R , ψ ∈ R ,
(4.4.8)
with R defined in Definition 4.2.1. The following stability property allows one to construct a solution
of (4.4.7):
Lemma 4.4.2 (A stability result). Let (uη )η be a sequence of uniformly Lipschitz continuous solutions of
the perturbed equation

ηu( y) + H̃ L ((0, z2 ), Du( y) + p2 e2 , y2 )
= λη in (−ρ, 0) × R,



R

ηu(
y)
+
H̃
((0,
z
),
Du(
y)
+
p
e
,
y
)

2
2 2
2
= λη in (0, ρ) × R,

 ηu( y) + max H̃ +,i ((0, z ), Dui ( y) + p e , y ) = λ
on Γ ,
2
2 2
2
η
i=L,R
(4.4.9)
−,L

ηu( y) + H̃ ((0, z2 ), Du( y) + p2 e2 , y2 )
= λη on {−ρ} × R,


 ηu( y) + H̃ −,R ((0, z ), Du( y) + p e , y )

= λη on {ρ} × R,
2
2 2
2


u is 1-periodic w.r.t. y2 ,
such that λη tends to λ as η tends to 0 and uη converges to u0 uniformly in [−ρ, ρ] × R. Then u0 is a
viscosity solution of (4.4.7) (replacing λρ (z2 , p2 ) with λ).
Proof. The proof of Lemma 4.4.2 follows the lines of the proofs of Theorem 6.1 and Theorem 6.2
in [AOT15b]. Actually, the proof is even simpler in the present case since the involved Hamiltonians do
not depend of η. We give it in Appendix 4.8.1 for the reader’s convenience.
The following comparison principle for (4.4.9) yields the uniqueness of the constant λρ (z2 , z2 ) for
which the cell-problem (4.4.7) admits a solution:
4.4. A NEW HAMILTONIAN INVOLVED IN THE EFFECTIVE TRANSMISSION CONDITION 137
Lemma 4.4.3 (A comparison result). For η > 0, let u be a bounded subsolution of (4.4.9) and v be a
bounded supersolution of (4.4.9). Then u ≤ v in [−ρ, ρ] × R.
Proof. As for Theorem 4.2.2, this result can be obtained by applying [Oud14, Theorem 3.4].
Lemma 4.4.4. There is a unique λρ (z2 , p2 ) ∈ R such that (4.4.7) admits a bounded solution. For this
choice of λρ (z2 , p2 ), there exists a solution χρ (z2 , p2 , ·) which is Lipschitz continuous with Lipschitz constant
L depending on p2 only (independent of ρ).
Proof. With the set M defined in (4.2.6), let us consider the new freezed dynamics fz2 : M → R2 and
running costs `z2 ,p2 : M → R2 :
fz2 ( y, a)
=
`z2 ,p2 ( y, a)
=
 
‹
1 −g 0 ( y2 ) L


f ((0, z2 ), a) if y1 ≤ 0, a ∈ AL ,
0
1

‹
1 −g 0 ( y2 ) R


f ((0, z2 ), a) if y1 ≥ 0, a ∈ AR ,
0
1
§ L
f2 ((0, z2 ), a)p2 + ` L ((0, z2 ), a) if y1 ≤ 0, a ∈ AL ,
f2R ((0, z2 ), a)p2 + `R ((0, z2 ), a) if y1 ≥ 0, a ∈ AR ,
(4.4.10)
(4.4.11)
where f2 stands for the second component of f .
Let Tz2 ,x,ρ be the set of admissible trajectories starting from y ∈ (−ρ, ρ)×R and constrained to [−ρ, ρ]×
R:


∞
+
ζ y ∈ Lip(R+ ; [−ρ, ρ] × R),

 (ζ y , a) ∈ Lloc (R ; M ) :
Z t
. (4.4.12)
Tz2 , y,ρ =
fz2 (ζ y (s), a(s))ds ∀t ∈ R+ 
ζ y (t) = y +

0
For any η > 0, the cost associated to the trajectory (ζ y , a) ∈ Tz2 , y,ρ is
Jρη (z2 , p2 ,
y; (ζ y , a)) =
Z
∞
`z2 ,p2 (ζ y (t), a(t))e−ηt d t,
(4.4.13)
Jρη (z2 , p2 , y; (ζ y , a)).
(4.4.14)
0
and we introduce the optimal control problem:
vρη (z2 , p2 , y) =
Thanks to [H3], we see that if δ00 =
inf
(ζ y ,a)∈Tz2 , y,ρ
δ0
p
,
2(1+kg 0 k∞ )
then B(0, δ00 ) ⊂ { fz2 ( y, a), a ∈ Ai } for any i = L, R,
e 3]" in §
y ∈ [−ρ, ρ] × R. This strong controlability property can be proved in the same manner as [H
0
4.2.2. From this, it follows that for any y, y ∈ [−ρ, ρ] × R,
|vρη (z2 , p2 , y) − vρη (z2 , p2 , y 0 )| ≤ L(p2 )| y − y 0 |.
(4.4.15)
for some L(p2 ) = L1 +L2 |p2 | with L1 , L2 depending on M f , M` δ0 and k g 0 k∞ but not on p2 . Introducing
χρη (z2 , p2 , y) = vρη (z2 , p2 , y) − vρη (z2 , p2 , (0, 0)), we deduce from (4.4.14) and (4.4.15) that there exists
a constant K = K(p2 ) such that
|ηvρη (z2 , p2 , y)| ≤ K,
and
|χρη (z2 , p2 , y)| ≤ K.
(4.4.16)
From Ascoli-Arzela’s theorem, up to the extraction a subsequence, χρη (z2 , p2 , ·) and −ηvρη (z2 , p2 , ·) converge uniformly respectively to a Lipschitz function χρ (z2 , p2 , ·) defined on [−ρ, ρ] × R (with Lipschitz
constant L) and to a constant λρ (z2 , p2 ) as η → 0.
On the other hand, with the arguments contained in [ACCT13, IM14a, IM14b], it can be proved that
vρη (z2 , p2 , ·) is a viscosity solution of (4.4.9) with λη = 0. Hence, χρη (z2 , p2 , ·) is a sequence of viscosity
solutions of (4.4.9) for λη = −ηvρη (z2 , p2 , (0, 0)), and λη → λρ (z2 , p2 ) as η tends to 0. From the stability
result in Lemma 4.4.2, the function χρ (z2 , p2 , ·) is a viscosity solution of (4.4.7). Finally, uniqueness
can be proved in a classical way using the comparison principle in Lemma 4.4.3 and the boundedness
of χρ .
138
4.4.2.2
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Passage to the limit as ρ → +∞
By definition of Tz2 , y,ρ , it is clear that if ρ1 ≤ ρ2 , then Tz2 , y,ρ1 ⊂ Tz2 , y,ρ2 . Then, thanks to (4.4.14) and
(4.4.16), we see that
−ηvρη ≤ −ηvρη ≤ K,
1
2
and letting η → 0, we obtain that
λρ1 (z2 , p2 ) ≤ λρ2 (z2 , p2 ) ≤ K.
(4.4.17)
Definition 4.4.1. We define the effective tangential Hamiltonian E(z2 , p2 ) as
E(z2 , p2 ) = lim λρ (z2 , p2 ).
(4.4.18)
ρ→∞
For z2 , p2 ∈ R fixed, we consider the global cell-problem

 H̃ i ((0, z2 ), Du( y) + p2 e2 , y2 )
max H̃ +,L ((0, z2 ), Du L ( y) + p2 e2 , y2 ), H̃ +,R ((0, z2 ), DuR ( y) + p2 e2 , y2 )
 u is 1-periodic w.r.t. y ,
2
= E(z2 , p2 )
= E(z2 , p2 )
in Ωi ,
on Γ ,
(4.4.19)
The following stability result is useful for proving the existence of a viscosity solution u of the cellproblem (4.4.19):
Lemma 4.4.5. Let uρ be a sequence of uniformly Lipschitz continuous solutions of the truncated cellproblem (4.4.7) which converges to u locally uniformly on R2 . Then u is a viscosity solution of the global
cell-problem (4.4.19).
Proof. Proceed exactly in the same way as in the proof of Lemma 4.4.2.
Theorem 4.4.1 (Existence of a global corrector). There exists χ(z2 , p2 , ·) a Lipschitz continuous viscosity
solution of (4.4.19) with the same Lipschitz constant L as in (4.4.15) and such that χ(z2 , p2 , (0, 0)) = 0.
Proof. Let χρ (z2 , p2 , ·) be the sequence of solutions of (4.4.7) given by Lemma 4.4.4. Recall that
χρ (z2 , p2 , ·) is Lipschitz continuous with Lipschitz constant L independent of ρ and periodic with
respect to y2 . By taking χρ (z2 , p2 , ·) − χρ (z2 , p2 , (0, 0)) instead of χρ (z2 , p2 , ·), we may assume that
χρ (z2 , p2 , (0, 0)) = 0. Thus, χρ (z2 , p2 , ·) is locally bounded and thanks to Ascoli-Arzela’s theorem, up to
the extraction a subsequence, χρ (z2 , p2 , ·) converges locally uniformy to a function χ(z2 , p2 , ·), which
is Lipschitz continuous and periodic with respect to y2 and satisfies χ(z2 , p2 , (0, 0)) = 0. Thanks to the
stability result in Lemma 4.4.5, χ(z2 , p2 , ·) is a viscosity solution of (4.4.19).
4.4.2.3
Comparison between E0 and E respectively defined in (4.3.9) and (4.4.18)
y
For " > 0, let us call W" (z2 , p2 , y) = "χ(z2 , p2 , " ). The following result is reminiscent of [GIM14,
Theorem 4.6,iii]:
Lemma 4.4.6. For any z2 , p2 ∈ R, there exists a subsequence "n such that W"n (z2 , p2 , ·) converges locally
uniformly to a Lipschitz function y 7→ W (z2 , p2 , y), with the Lipschitz constant L appearing in (4.4.15).
This function is constant with respect to y2 and satisfies W (z2 , p2 , 0) = 0. It is a viscosity solution of
H i ((0, z2 ), Du( y) + p2 e2 ) = E(z2 , p2 )
in Ωi .
(4.4.20)
Proof. It is clear that y 7→ W" (z2 , p2 , y) is a Lipschitz continuous function with constant L and that
W" (z2 , p2 , (0, 0)) = 0. Thus, from Ascoli-Arzela’s Theorem, we may assume that y 7→ W" (z2 , p2 , y) converges locally uniformly to some function y 7→ W (z2 , p2 , y), up to the extraction of subsequences. The
function y 7→ W (z2 , p2 , y) is Lipschitz continuous with constant L and W (z2 , p2 , (0, 0)) = 0. Moreover,
since W" (z2 , p2 , y) is periodic with respect to y2 with period ", W (z2 , p2 , y) does not depend on y2 .
To prove that W (z2 , p2 , ·) is a viscosity solution of (4.4.20), we first observe that y 7→ W" (z2 , p2 , ·) is a
viscosity solution of
y
H̃ i ((0, z2 ), Du( y) + p2 e2 , 2 ) = E(z2 , p2 ) in Ωi .
(4.4.21)
"
4.4. A NEW HAMILTONIAN INVOLVED IN THE EFFECTIVE TRANSMISSION CONDITION 139
For i = L, R, assume that ȳ ∈ Ωi , φ ∈ C 1 (Ωi ) and r0 < 0 are such that B( ȳ, r0 ) ⊂ Ωi and that
W (z2 , p2 , y) − φ( y) < W (z2 , p2 , ȳ) − φ( ȳ) = 0 for y ∈ B( ȳ, r0 ) \ { ȳ}.
We wish to prove that
H i ((0, z2 ), Dφ( ȳ) + p2 e2 ) ≤ E(z2 , p2 ).
Let us argue by contradiction and assume that there exists θ > 0 such that
H i ((0, z2 ), Dφ( ȳ) + p2 e2 ) = E(z2 , p2 ) + θ .
(4.4.22)
y2
" )
− δ, where δ > 0 is a fixed positive number. We claim that for
Take φ" ( y) = φ( y) + "∂ y1 φ( ȳ)g(
" > 0 and r > 0 small enough, φ" is a viscosity supersolution of
H̃ i ((0, z2 ), Du( y) + p2 e2 ,
y2
θ
) ≥ E(z2 , p2 ) +
"
2
in B( ȳ, r).
(4.4.23)
Indeed φ" is a regular function which satisfies
y
y
H̃ i ((0, z2 ), Dφ" ( y) + p2 e2 , 2 ) = H i (0, z2 ), Dφ( y) + g 0 ( 2 )(∂ y1 φ( ȳ) − ∂ y1 φ( y))e2 ,
"
"
and we deduce (4.4.23) from (4.4.22) and the regularity properties of the Hamiltonian H i . Hence,
W" (z2 , p2 , ·) is a subsolution of (4.4.21) and φ" is a supersolution of (4.4.23) in B( ȳ, r). Moreover
for r > 0 small enough, max y∈∂ B( ȳ,r) (W (z2 , p2 , y) − φ( y)) < 0. Hence, for " > 0 small enough
max y∈∂ B( ȳ,r) (W" (z2 , p2 , y) − φ" ( y)) ≤ 0.
Thanks to a standard comparison principle (which holds thanks to the fact that θ2 > 0)
max (W" (z2 , p2 , y) − φ" ( y)) ≤ 0.
y∈B( ȳ,r)
(4.4.24)
Letting " → 0 in (4.4.24), we deduce that
W (z2 , p2 , ȳ) ≤ φ( ȳ) − δ,
which is in contradiction with the assumption.
Using Lemma 4.4.6, it is possible to compare E0 (z2 , p2 ) and E(z2 , p2 ) respectively defined in (4.3.9)
and (4.4.18):
Proposition 4.4.1. For any z2 , p2 ∈ R,
E(z2 , p2 ) ≥ E0 (z2 , p2 ).
(4.4.25)
Proof. Let i ∈ {L, R} be fixed. Thanks to Lemma 4.4.6, the function y 7→ W (z2 , p2 , y) is a viscosity
solution of (4.4.20) in Ωi . Therefore, (4.4.20) is satisfied by W (z2 , p2 , ·) almost everywhere. Keeping in mind that W (z2 , p2 , y) is independent of y2 , we see that for almost all y ∈ Ωi , E(z2 , p2 ) =
H i ((0, z2 ), ∂ y1 W (z2 , p2 , y1 )e1 + p2 e2 ) ≥ E0i (z2 , p2 ).
From Proposition 4.4.1 and the coercivity of the Hamiltonian H i , the following numbers are well
defined for all z2 , p2 ∈ R:
L
Π (z2 , p2 )=max p ∈ R : H L ((0, z2 ), p2 e2 + pe1 ) = H −,L ((0, z2 ), p2 e2 + pe1 ) = E(z2 , p2 ) (4.4.26)
b L (z2 , p2 )=min p ∈ R : H L ((0, z2 ), p2 e2 + pe1 ) = H −,L ((0, z2 ), p2 e2 + pe1 ) = E(z2 , p2 ) (4.4.27)
Π
R
Π (z2 , p2 )=min p ∈ R : H R ((0, z2 ), p2 e2 + pe1 ) = H −,R ((0, z2 ), p2 e2 + pe1 ) = E(z2 , p2 ) (4.4.28)
b R (z2 , p2 )=max p ∈ R : H R ((0, z2 ), p2 e2 + pe1 ) = H −,R ((0, z2 ), p2 e2 + pe1 ) = E(z2 , p2 ) (4.4.29)
Π
Remark 4.4.3. In § 4.5, see in particular Corollaries 4.5.1 and 4.5.2 below, we will see that the function W
which is defined in Lemma 4.4.6 and provides information on the growth of y 7→ χ(z2 , p2 , y) as | y1 | → ∞,
satisfies
L
R
b L (z2 , p2 ) y1 1Ω L + Π
b R (z2 , p2 ) y1 1ΩR .
Π (z2 , p2 ) y1 1Ω L + Π (z2 , p2 ) y1 1ΩR ≤ W (z2 , p2 , y) ≤ Π
These growth properties at infinity show that χ(z2 , p2 , ·) is precisely the corrector associated to the reduced
set of test-functions proposed by Imbert and Monneau in [IM14a, IM14b], see § 4.7.1 below.
140
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Remark 4.4.4. From the convexity of the Hamiltonians H i and H −,i , we deduce that if E0i (z2 , p2 ) <
i
b i (z2 , p2 ). In this case, we will use the notation
E(z2 , p2 ), then Π (z2 , p2 ) = Π
i
b i (z2 , p2 ).
Πi (z2 , p2 ) = Π (z2 , p2 ) = Π
4.5
(4.4.30)
Further properties of the correctors
In this section, we prove further growth properties of the correctors, which will be useful in the proof of
convergence in § 4.7 below, see Remark 4.4.3. We start by stating a useful comparison principle related
to a mixed boundary value problem:
Lemma 4.5.1. Take 0 < ρ1 < ρ2 , z2 , p2 , λ ∈ R, a continuous function U0 : R → R and "0 > 0. Let v be a
continuous viscosity supersolution of
 R
H̃ ((0, z2 ), Dv( y) + p2 e2 , y2 ) ≥ λ,

 −,R
H̃ ((0, z2 ), Dv( y) + p2 e2 , y2 ) ≥ λ,
v( y) ≥ U0 ( y2 ),


v is 1-periodic w.r.t. y2 ,
y = ( y1 , y2 ) ∈ (ρ1 , ρ2 ) × R,
y = ( y1 , y2 ) ∈ {ρ2 } × R,
y = ( y1 , y2 ) ∈ {ρ1 } × R,
(4.5.1)
and u be a continuous viscosity subsolution of
 R
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 ) ≤ λ − "0 ,

 −,R
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 ) ≤ λ − "0 ,
u( y) ≤ U0 ( y2 ),


u is 1-periodic w.r.t. y2 ,
y = ( y1 , y2 ) ∈ (ρ1 , ρ2 ) × R,
y = ( y1 , y2 ) ∈ {ρ2 } × R,
y = ( y1 , y2 ) ∈ {ρ1 } × R,
(4.5.2)
where the inequalities on y1 = ρ1 are understood pointwise. Then, u ≤ v in [ρ1 , ρ2 ] × R.
Proof. The proof is rather classical, see [BCD97]. For the convenience of the reader, it is written in
Appendix 4.8.2.
Remark 4.5.1. From [IM14a, Proposition 2.14], we know that a bounded lsc function v is a supersolution
of (4.5.1) if and only if it is a supersolution of

 H̃ R ((0, z2 ), Dv( y) + p2 e2 , y2 ) ≥ λ, y = ( y1 , y2 ) ∈ (ρ1 , ρ2 ] × R,
v( y) ≥ U0 ( y2 ),
y = ( y1 , y2 ) ∈ {ρ1 } × R,
 v is 1-periodic w.r.t. y ,
2
and that a bounded usc function u is a subsolution of (4.5.2) if and only if it is a subsolution of

 H̃ R ((0, z2 ), Du( y) + p2 e2 , y2 ) ≤ λ − "0 , y = ( y1 , y2 ) ∈ (ρ1 , ρ2 ) × R,
u( y) ≤ U0 ( y2 ),
y = ( y1 , y2 ) ∈ {ρ1 } × R,
 u is 1-periodic w.r.t. y .
2
In other words, the boundary conditions on y1 = ρ2 correspond to state constraints.
Remark 4.5.2. There is of course a similar result for the mirror boundary value problem posed in [−ρ2 , −ρ1 ]×
R ⊂ Ω L with the Hamiltonian H̃ L instead of H̃ R , a Dirichlet condition on y1 = −ρ1 and a state constrained
boundary condition on y1 = −ρ2 (i.e. involving H̃ −,L ).
Proposition 4.5.1 (Control of slopes on the truncated domain). With E and E0R respectively defined
in (4.4.18) and (4.3.8), let z2 , p2 ∈ R be such that E(z2 , p2 ) > E0R (z2 , p2 ). There exists ρ ∗ = ρ ∗ (z2 , p2 ) > 0,
δ∗ = δ∗ (z2 , p2 ) > 0, m(·) = m(z2 , p2 , ·) : R+ → R+ satisfying limδ→0+ m(δ) = 0 and M ∗ = M ∗ (z2 , p2 )
such that for all δ ∈ (0, δ∗ ], ρ ≥ ρ ∗ , y = ( y1 , y2 ) ∈ [ρ ∗ , ρ] × R, h1 ∈ [0, ρ − y1 ] and h2 ∈ R,
χρ (z2 , p2 , y + h1 e1 + h2 e2 ) − χρ (z2 , p2 , y) ≥ (ΠR (z2 , p2 ) − m(δ))h1 − M ∗ ,
(4.5.3)
where ΠR (z2 , p2 ) is given by (4.4.30) and χρ (z2 , p2 , ·) is a solution of (4.4.7) given by Lemma 4.4.4.
4.5. FURTHER PROPERTIES OF THE CORRECTORS
141
Similarly, let z2 , p2 ∈ R be such that E(z2 , p2 ) > E0L (z2 , p2 ). There exists ρ ∗ > 0, δ∗ > 0, m(·) : R+ →
R+ and M ∗ as above, such that for all δ ∈ (0, δ∗ ], ρ ≥ ρ ∗ , y = ( y1 , y2 ) ∈ [−ρ, −ρ ∗ ] × R, h1 ∈ [0, ρ + y1 ]
and h2 ∈ R,
χρ (z2 , p2 , y − h1 e1 + h2 e2 ) − χρ (z2 , p2 , y) ≥ −(Π L (z2 , p2 ) + m(δ))h1 − M ∗ .
(4.5.4)
Proof. Let us focus on (4.5.3) since the proof of (4.5.4) is similar. Recall that ρ 7→ λρ (z2 , p2 ) is nondecreasing and tends to E(z2 , p2 ) as ρ → +∞. Choose ρ ∗ = ρ ∗ (z2 , p2 ) > 0 such that for any ρ ≥ ρ ∗ ,
E(z2 , p2 ) > λρ (z2 , p2 ) > E0R (z2 , p2 ). Then, choose δ∗ = δ∗ (z2 , p2 ) > 0 such that for any δ ∈ (0, δ∗ ],
λρ (z2 , p2 ) − δ > E0R (z2 , p2 ).
q 7→ H −,R ((0, z2 ), p2 e2 + qe1 )
E(z2 , p2 )
λρ∗ (z2 , p2 )
λρ (z2 , p2 ) − δ
λρ∗ (z2 , p2 ) − δ∗
E0R (z2 , p2 )
qδ ∗
qδ
ΠR (z2 , p2 )
Figure 4.2: Construction of ρ ∗ , δ∗ and qδ
Let us fix ρ > ρ ∗ , δ ∈ (0, δ∗ ] and ȳ = ( ȳ1 , ȳ2 ) ∈ [ρ ∗ , ρ] × R. Consider y 7→ χρ (z2 , p2 , y) a solution
of (4.4.7) as in Lemma 4.4.4. The function χρ (z2 , p2 , ·) is 1-periodic with respect to y2 and Lipschitz
continuous with constant L = L(p2 ). Thus, for any y ∈ { ȳ1 } × R
χρ (z2 , p2 , y) − χρ (z2 , p2 , ȳ) ≥ −L.
Therefore, y 7→ χρ (z2 , p2 , y) − χρ (z2 , p2 , ȳ) is a supersolution of
 R
H̃ ((0, z2 ), Dv( y) + p2 e2 , y2 ) ≥ λρ (z2 , p2 ),

 −,R
H̃ ((0, z2 ), Dv( y) + p2 e2 , y2 ) ≥ λρ (z2 , p2 ),
v( y) ≥ −L,


v is 1-periodic w.r.t. y2 .
y ∈ ( ȳ1 , ρ) × R,
y ∈ {ρ} × R,
y ∈ { ȳ1 } × R,
(4.5.5)
Since ρ ≥ ρ ∗ and δ ∈ (0, δ∗ ], there exists a unique qδ ∈ R, see Figure 4.2, such that
λρ (z2 , p2 ) − δ = H R ((0, z2 ), p2 e2 + qδ e1 ) = H −,R ((0, z2 ), p2 e2 + qδ e1 ).
(4.5.6)
We observe that qδ∗ ≤ qδ ≤ ΠR (z2 , p2 ) and that qδ tends to ΠR (z2 , p2 ) as δ tends to 0. Choose m(δ) =
ΠR (z2 , p2 ) − qδ ≥ 0 and consider the function wR : wR ( y) = qδ ( y1 + g( y2 )), which is of class C 2 . From
the choice of qδ , for any y ∈ R2 ,
H̃ −,R (z2 , DwR ( y) + p2 e2 , y2 ) ≤ H̃ R (z2 , DwR ( y) + p2 e2 , y2 ) = λρ (z2 , p2 ) − δ,
(4.5.7)
and for any y ∈ { ȳ1 } × R,
wR ( y) − qδ ȳ1 ≤ |qδ | k g k∞ ≤ max |qδ∗ |, |ΠR (z2 , p2 )| k g k∞ = C = C(z2 , p2 ).
Therefore, as stated in Remark 4.5.1, the subsolution property holds up to the boundary y1 = ρ and
the function uR : y ∈ [ ȳ1 , R] × R 7→ wR ( y) − qδ ȳ1 − C − L is a subsolution of
 R
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 ) ≤ λρ (z2 , p2 ) − δ
y ∈ ( ȳ1 , ρ) × R,

 −,R
H̃ ((0, z2 ), Du( y) + p2 e2 , y2 ) ≤ λρ (z2 , p2 ) − δ, y ∈ {ρ} × R,
(4.5.8)
u( y) ≤ −L,
y ∈ { ȳ1 } × R,


u is 1-periodic w.r.t. y2 .
142
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Finally, since χρ (z2 , p2 , y) − χρ (z2 , p2 , ȳ) is a supersolution of (4.5.5) and uR is a subsolution of (4.5.8),
the comparison principle in Lemma 4.5.1 yields: for all y ∈ [ ȳ1 , ρ] × R
χρ (z2 , p2 , y) − χρ (z2 , p2 , ȳ) ≥ uR ( y)
=
≥
qδ (( y1 − ȳ1 ) + g( y2 )) − C − L
(ΠR (z2 , p2 ) − m(δ))( y1 − ȳ1 ) − M ∗ ,
(4.5.9)
where M ∗ is a constant depending only of z2 and p2 . Note that the constants which appear in (4.5.9)
are independent of ρ > 0.
The following corollary deals with the global corrector χ:
Corollary 4.5.1. With Πi (z2 , p2 ) defined in (4.4.30), i = L, R,
1. If E(z2 , p2 ) > E0R (z2 , p2 ), then, with ρ ∗ > 0 and M ∗ ∈ R as in the first point of Proposition 4.5.1, for
all y ∈ [ρ ∗ , +∞) × R, h1 ≥ 0 and h2 ∈ R,
χ(z2 , p2 , y + h1 e1 + h2 e2 ) − χ(z2 , p2 , y) ≥ ΠR (z2 , p2 )h1 − M ∗ .
(4.5.10)
2. If E(z2 , p2 ) > E0L (z2 , p2 ), then, with ρ ∗ > 0 and M ∗ ∈ R as in the second point of Proposition 4.5.1,
for all y ∈ (−∞, −ρ ∗ ] × R, h1 ≥ 0 and h2 ∈ R,
χ(z2 , p2 , y − h1 e1 + h2 e2 ) − χ(z2 , p2 , y) ≥ −Π L (z2 , p2 )h1 − M ∗ .
(4.5.11)
Proof. The proof follows easily from Proposition 4.5.1 and the local uniform convergence of the sequence χρ (z2 , p2 , ·) toward χ(z2 , p2 , ·).
Corollary 4.5.2. For any z2 , p2 ∈ R, the function y 7→ W (z2 , p2 , y) defined in Lemma 4.4.6 satisfies
R
b R (z2 , p2 )
Π (z2 , p2 ) ≤ ∂ y1 W (z2 , p2 , y) ≤ Π
L
Π (z2 , p2 ) ≤ ∂ y1 W (z2 , p2 , y) ≤ Π (z2 , p2 )
bL
for a.a. y ∈ (0, +∞) × R,
(4.5.12)
for a.a. y ∈ (−∞, 0) × R,
(4.5.13)
and for all y:
L
R
b R (z2 , p2 ) y1 1ΩR .
b L (z2 , p2 ) y1 1Ω L + Π
Π (z2 , p2 ) y1 1Ω L + Π (z2 , p2 ) y1 1ΩR ≤ W (z2 , p2 , y) ≤ Π
(4.5.14)
Proof. From Lemma 4.4.6, we see that y 7→ W (z2 , p2 , y) is Lipschitz continuous w.r.t. y1 and independent of y2 , and satisfies
H R ((0, z2 ), ∂ y1 W (z2 , p2 , y)e1 + p2 e2 ) = E(z2 , p2 )
a.e. in ΩR .
(4.5.15)
Consider first the case when E(z2 , p2 ) > E0R (z2 , p2 ); from the convexity and coercivity of H R , the observations above yield that almost everywhere in y, ∂ y1 W (z2 , p2 , y) can be either ΠR (z2 , p2 ) (the unique
real number such that H −,R ((0, z2 ), qe1 + p2 e2 ) = E(z2 , p2 )), or the unique real number q (depending
on (z2 , p2 )) such that H +,R ((0, z2 ), qe1 + p2 e2 ) = E(z2 , p2 ). Note that q < ΠR (z2 , p2 ). But from Corollary
4.5.1 and the local uniform convergence of W" (z2 , p2 , ·) toward W (z2 , p2 , ·), we see that that for any
y1 > 0 and h1 ≥ 0,
W (z2 , p2 , y + h1 e1 ) − W (z2 , p2 , y) ≥ ΠR (z2 , p2 )h1 ,
which implies that almost everywhere, ∂ y1 W (z2 , p2 , y) ≥ ΠR (z2 , p2 ) > q. Therefore, ∂ y1 W (z2 , p2 , ·) =
ΠR (z2 , p2 ) almost everywhere.
R
In the case when E(z2 , p2 ) = E0R (z2 , p2 ), we deduce from (4.5.15) that almost everywhere in y, Π (z2 , p2 ) ≤
b R (z2 , p2 ).
∂ y1 W (z2 , p2 , y) ≤ Π
We have proved (4.5.12). The proof of (4.5.13) is identical. Finally, (4.5.14) comes from (4.5.12), (4.5.13)
and from the fact that W (z2 , p2 , 0) = 0.
4.6
A comparison principle for the effective equation
To prove the main result of the paper, i.e. Theorem 4.1.2, we need a comparison principle for (4.1.21)(4.1.22). Before proving such a result, we need to establish some useful properties of E arising in
(4.1.22).
4.6. A COMPARISON PRINCIPLE FOR THE EFFECTIVE EQUATION
4.6.1
143
Properties of E(·, ·)
In the theory of homogenization of Hamilton-Jacobi equations, it is quite standard to observe that
the effective Hamiltonian inherits some properties from the original problem, see the pioneering work
[LPV].
Lemma 4.6.1. For any z2 ∈ R, the function p2 7→ E(z2 , p2 ) is convex. For any z2 , z20 , p2 , p20 ∈ R,
| E(z2 , p2 ) − E(z2 , p20 ) |≤ M f |p2 − p20 |,
| E(z2 , p2 ) −
E(z20 , p2 )
(4.6.1)
− z20 |),
(4.6.2)
δ0 |p2 | − M` ≤ E(z2 , p2 ) ≤ M f |p2 | + M` ,
(4.6.3)
|≤
C(1 + |p2 |)|z2 − z20 | + ω(|z2
where the constants M f , M` , δ0 have been introduced in Assumptions [H0]-[H3], the modulus of continuity
ω has been introduced in Lemma 4.4.1 and C is a positive constant.
Moreover, p2 7→ E(z2 , p2 ) is affine in a neighborhood of ±∞. More precisely, for any z2 ∈ R, there exist
ˆ`(z2 ), ˇ`(z2 ) ∈ [−M` , M` ], fˆ(z2 ), fˇ(z2 ) ∈ [δ0 , M f ] and K̂(z2 ), Ǩ(z2 ) > 0 such
p2 ≥ K̂(z2 )
⇒
p2 ≤ −Ǩ(z2 )
⇒
E(z2 , p2 ) = fˆ(z2 )p2 + ˆ`(z2 ),
E(z2 , p2 ) = − fˇ(z2 )p2 + ˇ`(z2 ).
(4.6.4)
(4.6.5)
Proof. The proof contains arguments that are quite similar to those contained in [LPV], but technical
difficulties arise from the discontinuities of the Hamiltonians at y1 = 0.
Proof of (4.6.1) For z2 , p2 , p20 ∈ R, consider y 7→ vρη (z2 , p2 , y) and y 7→ vρη (z2 , p20 , y) given by (4.4.14).
These functions are the viscosity solutions of (4.4.9) with λη = 0. Take ȳ ∈ R2 and ϕ ∈ Rρ (see
M
(4.4.8)) and assume that y 7→ vρη (z2 , p20 , y) + ηf |p2 − p20 | − ϕ( y) has a local minimum at ȳ. We focus on the case when ȳ = (0, ȳ2 ) ∈ Γ , since all the other cases are simpler. Therefore ηvρη (z2 , p20 , ȳ) +
maxi∈{L,R} H̃ +,i ((0, z2 ), Dϕ i ( ȳ) + p20 e2 , ȳ2 ) ≥ 0. Assume for instance that the latter maximum is achieved
by i = R, then
€
Š
M
η vρη (z2 , p20 , ȳ) + ηf |p2 − p20 | + H̃ +,R ((0, z2 ), Dϕ R ( ȳ) + p2 e2 , ȳ2 ) ≥ M f |p2 − p20 |
+H̃ +,R ((0, z2 ), Dϕ R ( ȳ) + p2 e2 , ȳ2 ) − H̃ +,R ((0, z2 ), Dϕ R ( ȳ) + p20 e2 , ȳ2 ),
and from (4.4.4)

‹
Mf
η vρη (z2 , p20 , ȳ) +
|p2 − p20 | + H̃ +,R ((0, z2 ), Dϕ R ( ȳ) + p2 e2 , ȳ2 ) ≥ 0.
η
This shows that y 7→ vρη (z2 , p20 , y) +
Mf
η
|p2 − p20 | is a supersolution of the equation satisfied by y 7→
M
vρη (z2 , p2 , y). The comparison principle in Lemma 4.4.3 yields vρη (z2 , p20 , ·) + ηf |p2 − p20 | ≥ vρη (z2 , p2 , ·)
and (4.6.1) is obtained by exchanging the roles of p2 and p20 and passing to the limit as η → 0 then as
ρ → +∞.
Proof of (4.6.2) For z2 , z20 , p2 ∈ R, consider y 7→ vρη (z2 , p2 , y) and y 7→ vρη (z20 , p2 , y) given by (4.4.14).
These functions are viscosity solutions of (4.4.9) with λη = 0. Assume that 0 = vρη (z20 , p2 , ȳ) − ϕ( ȳ) be
a local minimum of vρη (z20 , p2 , ·) − ϕ(·) for ȳ ∈ R2 and ϕ ∈ Rρ . As above, we focus on the case when
ȳ ∈ Γ because the other cases are simpler. It is not restrictive to assume that ϕ L and ϕ R are smooth
(at least C 3 ). Since y 7→ vρη (z20 , p2 , y) is Lipschitz continuous with a constant L(p2 ) = L1 + L2 |p2 |, see
(4.4.15), we see that
|∂ y2 ϕ L ( ȳ)| = |∂ y2 ϕ R ( ȳ)| ≤ L(p2 ),
∂ y1 ϕ L ( ȳ) ≥ −L(p2 ),
∂ y1 ϕ R ( ȳ) ≤ L(p2 ).
(4.6.6)
It is always possible to modify ϕ and obtain a test-function ψ such that |∂ y1 ψi ( ȳ)| ≤ 2L(p2 ), |∂ y2 ψi ( ȳ)| ≤
L(p2 ), i = L, R and that vρη (z20 , p2 , ·) − ψ(·) has a local minimum at ȳ. Indeed, we make out two cases:
1. if |∂ y1 ϕ i ( ȳ)| ≤ 2L(p2 ) for i = L, R, then we choose ψ = ϕ.
144
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
2. If ∂ y1 ϕ L ( ȳ) > 2L(p2 ) or ∂ yw 1 ϕ R ( ȳ) < −2L(p2 ), let us introduce

ϕ( y) − (2L(p2 ) + ∂ y1 ϕ R ( ȳ)) y1 − A| y − ȳ|2


ϕ( y) − A| y − ȳ|2
ψ( y) =
ϕ( y) + (2L(p2 ) − ∂ y1 ϕ L ( ȳ)) y1 − A| y − ȳ|2


ϕ( y) − A| y − ȳ|2
if
if
if
if
y
y
y
y
∈ [0, ρ] × R and ∂ y1 ϕ R ( ȳ) < −2L(p2 ),
∈ [0, ρ] × R and |∂ y1 ϕ R ( ȳ)| ≤ 2L(p2 ),
∈ [−ρ, 0] × R and ∂ y1 ϕ L ( ȳ) > 2L(p2 ),
∈ [−ρ, 0] × R and |∂ y1 ϕ L ( ȳ)| ≤ 2L(p2 ).
Note that |∂ y1 ψi ( ȳ)| ≤ 2L(p2 ), i = L, R and that
∂ y1 ϕ L ( ȳ) ≥ ∂ y1 ψ L ( ȳ),
and
∂ y1 ϕ R ( ȳ) ≤ ∂ y1 ψR ( ȳ).
(4.6.7)
We claim that for A large enough, vρη (z20 , p2 , ·) − ψ(·) has a local minimum at ȳ. Indeed, fix
r > 0 such that vρη (z20 , p2 , y) − ϕ( y) ≥ vρη (z20 , p2 , ȳ) − ϕ( ȳ) = 0 for y ∈ B( ȳ, r). Assuming for
example that ∂ y1 ϕ R ( ȳ) < −2L(p2 ) (the case when |∂ y1 ϕ R ( ȳ)| ≤ 2L(p2 ) is obvious), we see that
for y ∈ B( ȳ, r) with y1 > 0,
vρη (z20 , p2 , y) − ψ( y) ≥vρη (z20 , p2 , y) − vρη (z20 , p2 , (0, y2 )) + ϕ(0, y2 ) − ϕ( y)
+ 2L(p2 ) + ∂ y1 ϕ R ( ȳ) y1 + A| y − ȳ|2 .
(4.6.8)
For a constant c > 0 depending on φ and r,
ϕ(0, y2 ) − ϕ( y)
≥
− y1 ∂ y1 ϕ R (0, y2 ) − c y12 ≥ − y1 ∂ y1 ϕ R ( ȳ) − c( y12 + y1 | y2 − ȳ2 |).
On the other hand, vρη (z20 , p2 , y)− vρη (z20 , p2 , (0, y2 )) ≥ −L(p2 ) y1 . From the latter two observations
and (4.6.8), we deduce that vρη (z20 , p2 , y) − ψ( y) ≥ L(p2 ) y1 − c( y12 + y1 | y2 − ȳ2 |) + A| y − ȳ|2 .
Therefore for A large enough,
vρη (z20 , p2 , y) − ψ( y) ≥ 0 = vρη (z20 , p2 , ȳ) − ψ( ȳ),
y ∈ B( ȳ, r),
and the claim is proved.
In both cases, we see that ηvρη (z20 , p2 , ȳ) + maxi∈{L,R} H̃ +,i ((0, z20 ), Dψi ( ȳ) + p2 e2 , ȳ2 ) ≥ 0; assuming
that the latter maximum is achieved by i = R for example, this yields
Š
€
η vρη (z20 , p2 , ȳ) + Cη (1 + |p2 |)|z2 − z20 | + η1 ω(|z2 − z20 |) + H̃ +,R ((0, z2 ), DψR ( ȳ) + p2 e2 , ȳ2 )
≥ C(1 + |p2 |)|z2 − z20 | − M |p2 e2 + DψR ( ȳ)||z2 − z20 |.
Thanks to (4.6.6) and from the construction of ψ,
|p2 e2 + Dψi ( ȳ)| ≤ |p2 | + 3L(p2 ),
i = L, R.
(4.6.9)
Hence, choosing C ≥ M max(3L1 , 1 + 3L2 ) yields
Š
€
η vρη (z20 , p2 , ȳ) + Cη (1 + |p2 |)|z2 − z20 | + η1 ω(|z2 − z20 |) + H̃ +,R ((0, z2 ), DψR ( ȳ) + p2 e2 , ȳ2 ) ≥ 0.
But from (4.6.7) and the nonincreasing character of H̃ +,R , we see that
€
Š
η vρη (z20 , p2 , ȳ) + Cη (1 + |p2 |)|z2 − z20 | + η1 ω(|z2 − z20 |) + H̃ +,R ((0, z2 ), Dϕ R ( ȳ) + p2 e2 , ȳ2 ) ≥ 0.
Therefore, vρη (z20 , p2 , ·) + Cη (1 + |p2 |)|z2 − z20 | + η1 ω(|z2 − z20 |) is a supersolution of the equation satisfied
by vρη (z2 , p2 , ·) and we conclude using the comparison principle Lemma 4.4.3, passing to the limit as
η → 0 and ρ → +∞, and finally exchanging the roles of z2 and z20 .
Proof of (4.6.3) Choosing ā ∈ Ai such that f i ((0, z2 ), ā) = −δ0 e2 if p2 ≥ 0 or f i ((0, z2 ), ā) = +δ0 e2
otherwise, (see [H3]), implies H i ((0, z2 ), p2 e2 + p1 e1 ) ≥ δ0 |p2 | − M` for all p1 ∈ R. This yields that
E0 (z2 , p2 ) ≥ δ0 |p2 | − M` , and the lower bound in (4.6.3) is obtained thanks to Proposition 4.4.1. The
upper bound is obtained by bounding −ηvρη by M` + M f |p2 | from above, and passing to the limit as
η → 0 and ρ → +∞.
4.6. A COMPARISON PRINCIPLE FOR THE EFFECTIVE EQUATION
145
Proof ot the convexity property It is well known that p2 7→ Jρη (z2 , p2 , y; (γ y , a)) (see (4.4.13)) is
an affine function. This implies that p2 7→ −ηvρη (z2 , p2 , y) is convex, (see (4.4.14) for the definition of
vρη (z2 , p2 , y)). The convexity of p2 7→ E(z2 , p2 ) is then obtained by letting η → 0 then ρ → ∞.
Proof that p2 7→ E(z2 , p2 ) is affine in a neighborhood of ±∞ We focus on (4.6.4) since the proof
of (4.6.5) is similar. For z2¦∈ R, y ∈ [−ρ, ρ] × R and p2©, η, ρ > 0, let us define
R∞
f¯ρη (z2 , y) = sup(γ y ,a)∈Tz , y,ρ − 0 f2 ((0, z2 ), a(t))e−ηt d t , with Tz2 , y,ρ given in (4.4.12). From (4.4.11)
2
and (4.4.14), we deduce that
η f¯ρη (z2 , y)p2 − M` ≤ −ηvρη (z2 , p2 , y) ≤ η f¯ρη (z2 , y)p2 + M` .
(4.6.10)
From the assumptions, it is easy to check that
|η f¯ρη (z2 , y)|
≤
| f¯ρη (z2 , y) − f¯ρη (z2 , y 0 )|
≤
Mf ,
(4.6.11)
0
C| y − y |,
(4.6.12)
for some positive constant C, and that y 7→ f¯ρη (z2 , y) is periodic with period 1 in the variable y2 . From
Ascoli-Arzela theorem, up to the extraction of subsequences, we may assume that f¯η (z2 , ·) − f¯η (z2 , 0)
ρ
ρ
tends to a Lipschitz function (with Lipschitz constant C) and that η f¯ρη (z2 , ·) tends to a constant f¯ρ (z2 )
as η → 0. With the same arguments as in § 4.4.2.2, we may prove that f¯ρ (z2 ) is nondecreasing and
uniformly bounded with respect to ρ. Therefore, we may define fˆ(z2 ) = limρ→+∞ f¯ρ (z2 ).
Passing to the limit in (4.6.10) as η → 0 then as ρ → +∞, we deduce that
fˆ(z2 )p2 − M` ≤ E(z2 , p2 ) ≤ fˆ(z2 )p2 + M` .
(4.6.13)
Finally, from (4.6.13) and the convexity of p2 7→ E(z2 , p2 ), we infer that there exists ˆ`(z2 ) ∈ [−M` , M` ]
and K̂(z2 ) > 0 such that for any p2 ≥ K̂(z2 )
E(z2 , p2 ) = fˆ(z2 )p2 + ˆ`(z2 ).
Finally, the bound fˆ(z2 ) ≥ δ0 comes from (4.6.3), and it is straightforward to check that fˆ(z2 ) ≤
Mf .
4.6.2
The comparison principles
Since we are not able to control the constants K̂(z2 ) and Ǩ(z2 ) arising in Lemma 4.6.1, we cannot
directly use the comparison principle which is available in [Oud14, Theorem 2.5]. To apply the latter,
it will be useful to first modify E(z2 , p2 ) for |p2 | larger than some fixed number K independent of z2 .
The following lemma deals with such modified Hamiltonians.
Lemma 4.6.2. For a positive number K, the Hamiltonian EK (z2 , p2 ) defined by

E(z2 , p2 ) if |p2 | ≤ K,

E(z2 , K) + M f (p2 − K) if p2 > K,
EK (z2 , p2 ) =
 E(z , −K) − M (p + K) if p < −K,
2
f
2
2
(4.6.14)
is convex in the variable p2 and
EK (z2 , p2 ) =
max
b∈[−M f ,M f ]
bp2 − EK? (z2 , b) ,
(4.6.15)
where EK? : R2 → R ∪ {+∞} is the Fenchel transform EK? (z2 , b) = supq∈R (bq − EK (z2 , q)). For z2 , z20 ∈ R
and b, b0 ∈ [−M f , M f ],
|EK? (z2 ,
b) −
EK? (z20 ,
|EK? (z2 ,
|EK? (z2 , b)| ≤ CK ,
b) −
b)| ≤ ωK (t) = C(1 +
EK? (z2 ,
0
0
b )| ≤ K|b − b |,
K)|z2 − z20 | + ω(|z2
− z20 |),
(4.6.16)
(4.6.17)
(4.6.18)
where in (4.6.16), CK is a positive constant, and, in (4.6.18), the constant C and the modulus of continuity
ω are those appearing in (4.6.2).
146
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Proof. The convexity of p2 7→ EK (z2 , p2 ) comes from the convexity of p2 7→ E(z2 , p2 ) and from (4.6.1).
From (4.6.1), it is also clear that EK? (z2 , b) = +∞ if b ∈
/ [−M f , M f ], which implies (4.6.15). It can also
be seen that if b ∈ [−M f , M f ], then
EK? (z2 , b) = max (bp − EK (z2 , p)) .
(4.6.19)
p∈[−K,K]
From (4.6.3), we check that for z2 ∈ R, p2 ∈ [−K, K] and b ∈ [−M f , M f ],
bp2 − M f |p2 | − M` ≤ bp2 − EK (z2 , p2 ) ≤ bp2 + M f (K − |p2 |) − δ0 K + M` .
(4.6.20)
Choosing p2 = 0 yields that EK? (z2 , b2 ) ≥ −M` . Using the fact that bp2 − M f |p2 | ≤ 0 if |b| ≤ M f , we
deduce that EK? (z2 , b) ≤ CK and finally (4.6.16) with CK = (M f − δ0 )K + M` .
It is standard to deduce (4.6.17) and (4.6.18) from (4.6.19) and (4.6.2).
Lemma 4.6.2 allows us to prove the following comparison principle:
Proposition 4.6.1. Let u and w be respectively a bounded subsolution and a bounded supersolution of
λu(z) + H i (z, Du(z)) = 0
λu(z) + max EK (z2 , ∂z2 ϕ(z)), HΓ (z, Du L (z), DuR (z)) = 0
if z ∈ Ωi ,
if z = (0, z2 ) ∈ Γ .
(4.6.21)
Then, u ≤ w in R2 .
Proof. By virtue of Lemma 4.6.2, it is possible to apply [Oud14, Theorem 2.5], more precisely the
general comparison principle discussed in [Oud14, Remark 2.11], since we do not assume the continuity
of u.
Theorem 4.6.1. Let u : R2 → R and v : R2 → R be respectively a bounded subsolution and a bounded
supersolution of (4.1.23). Then u ≤ v in R2 .
Proof. Since u is a subsolution of (4.1.23), it is also a subsolution of
λu(z) + H i (z, Du(z)) = 0
λu(z) + HΓ (z, Du L (z), DuR (z)) = 0
if z ∈ Ωi ,
if z = (0, z2 ) ∈ Γ .
Thanks to Assumptions [H0]-[H3], we can apply [Oud14, Lemma 2.6] to u: there exists r > 0 such
that u is Lipschitz continuous in [−r, r] × R. Let us call Lu the Lipschitz constant of the restriction of
u to [−r, r] × R and choose K ≥ Lu . Since EK coincides with E on R × [−K, K], we deduce that u is a
subsolution of (4.6.21).
On the other hand, since EK ≥ E, v is a supersolution of (4.6.21).
The proof is achieved by applying Proposition 4.6.1 to the pair (u, v).
4.7
4.7.1
Proof of the main result
A reduced set of test-functions
From [IM14a] and [IM14b], we may use an equivalent definition for the viscosity solution of (4.1.23).
i
b i , i = L, R, have been introduced in (4.4.26)-(4.4.29). Let Π :
Definition 4.7.1. Recall that Π and Π
2
2
L
R
R → R , (z2 , p2 ) 7→ Π (z2 , p2 ), Π (z2 , p2 ) be such that, for all (z2 , p2 )
b L (z2 , p2 ) ≤ Π L (z2 , p2 ) ≤ Π L (z2 , p2 ),
Π
R
b R (z2 , p2 ).
Π (z2 , p2 ) ≤ ΠR (z2 , p2 ) ≤ Π
(4.7.1)
For z̄ = (0, z̄2 ) ∈ Γ , the reduced set of test-functions R Π (z̄) associated to the map Π is the set of the functions
ϕ ∈ C 0 (R2 ) such that there exists a C 1 function ψ : R → R with
ϕ(z) = ψ(z2 ) + ΠR z̄2 , ψ0 (z̄2 ) 1z∈ΩR + Π L z̄2 , ψ0 (z̄2 ) 1z∈Ω L z1 .
(4.7.2)
4.7. PROOF OF THE MAIN RESULT
147
The following theorem is reminiscent of [IM14a, Theorem 2.6].
Theorem 4.7.1. Let u : R2 → R be asubsolution (resp. supersolution) of (4.1.21) and a map Π : R2 →
R2 , (z2 , p2 ) 7→ Π L (z2 , p2 ), ΠR (z2 , p2 ) such that (4.7.1) holds for all (z2 , p2 ) ∈ R2 .
The function u is a subsolution (resp. supersolution) of (4.1.22) if and only if for any z = (0, z2 ) ∈ Γ and
for all ϕ ∈ R Π (z) such that u − ϕ has a local maximum (resp. local minimum) at z,
λu(z) + max E(z2 , ∂z2 ϕ(z)), HΓ (z, Dϕ L (z), Dϕ R (z)) ≤ 0, (resp. ≥ 0),
(4.7.3)
where the meaning of Dϕ L and Dϕ R is given in Definition 4.1.1.
Proof. The proof follows the lines of that of [IM14a, Theorem 2.6] and is given in Appendix 4.8.3 for the
reader’s convenience. It is worth to note that Lemma 4.3.1 is important in order to use the arguments
contained in the proof of [IM14a, Theorem 2.6].
4.7.2
Proof of Theorem 4.1.2
As seen in § 4.3, the result will be proved if we show that the sequence (ṽ" )" corresponding to the
straightened geometry converges to v. We will actually prove that ṽ and ṽ defined in (4.3.1) are respectively a subsolution and a supersolution of (4.1.23). From Theorem 4.6.1, this will imply that
ṽ = ṽ = v = lim"→0 ṽ" . Moreover, from Proposition 4.3.1, we just have to check the transmission
condition (4.1.22).
We restrict ourselves to checking that ṽ is a subsolution of (4.1.23), since the proof that ṽ is a
supersolution of (4.1.23) is similar. Take z̄ = (0, z̄2 ) ∈ Γ€. We are going to use
Š Theorem 4.7.1 with
L
R
the special choice for the map Π : R2 → R2 : Π(z2 , p2 ) = Π (z2 , p2 ), Π (z2 , p2 ) . Take a test-function
ϕ ∈ R Π (z̄), i.e. of the form
Š
€ R
L
ϕ(z) = ψ(z2 ) + Π (z̄2 , ψ0 (z̄2 ))1z∈ΩR + Π (z̄2 , ψ0 (z̄2 ))1z∈Ω L z1 ,
(4.7.4)
for a C 1 function ψ : R → R, such that ṽ − ϕ has a strict local maximum at z̄ and that ṽ(z̄) = ϕ(z̄).
Let us proceed by contradiction and assume that
(4.7.5)
λϕ(z̄) + max E(z̄2 , ∂z2 ϕ(z̄)), HΓ (z̄, Dϕ L (z̄), Dϕ R (z̄)) = θ > 0.
From (4.7.4), we see that HΓ (z̄, Dϕ L (z̄), Dϕ R (z̄)) ≤ E(z̄2 , ∂z2 ϕ(z̄)) and (4.7.5) is equivalent to
λψ(z̄2 ) + E(z̄2 , ψ0 (z̄2 )) = θ > 0.
(4.7.6)
Let χ(z̄2 , ψ0 (z̄2 ), ·) be a solution of (4.4.19) such that χ(z̄2 , ψ0 (z̄2 ), (0, 0)) = 0, see Theorem 4.4.1, and
W (z̄2 , ψ0 (z̄2 ), z1 ) = lim"→0 "χ(z̄2 , ψ0 (z̄2 ), "z ).
We claim that for " > 0 and r > 0 small enough, the function ϕ " :
Step 1
z
ϕ " (z) = ψ(z2 ) + "χ(z̄2 , ψ0 (z̄2 ), )
"
is a viscosity supersolution of
λϕ " (z) + H̃"i (z, Dϕ " (z))
"
λϕ (z) + H̃Γ ," (z, D (ϕ " ) L (z), D (ϕ " )R (z))
≥
≥
θ
2
θ
2
if z ∈ Ωi ∩ B((0, z̄2 ), r),
if z ∈ Γ ∩ B((0, z̄2 ), r),
(4.7.7)
where the Hamiltonians H̃"i and H̃Γ ," are defined by (4.2.11) and (4.2.12).
Indeed, if ξ is a test-function in R such that ϕ " − ξ has a local minimum at z ? ∈ B((0, z̄2 ), r), then, from
?
the definition of ϕ " , y 7→ χ(z̄2 , ψ0 (z̄2 ), y) − 1" (ξ(" y) − ψ(" y2 )) has a local minimum at z" . Let us now
0
use the fact that y 7→ χ(z̄2 , ψ (z̄2 ), y) is a supersolution of (4.4.19):
?
If z" ∈ Ωi , for i = L or R, then H̃ i ((0, z̄2 ), Dξ(z ? ) − ψ0 (z2? )e2 + ψ0 (z̄2 )e2 ,
regularity properties of H̃ i , see Lemma 4.4.1,
H̃ i ((0, z̄2 ), Dξ(z ? ) − ψ0 (z2? )e2 + ψ0 (z̄2 )e2 ,
z2?
"
z2?
" )
≥ E(z̄2 , ψ0 (z̄2 )). From the
) = H̃"i (z ? , Dξ(z ? )) + o"→0 (1) + o r→0 (1),
148
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
thus
λϕ " (z ? ) + H̃"i (z ? , Dξ(z ? ))
‹‹


z?
+ o"→0 (1) + o r→0 (1).
≥E(z̄2 , ψ0 (z̄2 )) + λ ψ(z2? ) + "χ z̄2 , ψ0 (z̄2 ),
"
From (4.7.6), this implies that
‹

z?
λϕ " (z ? ) + H̃"i (z ? , Dξ(z ? )) ≥ θ + λ"χ z̄2 , ψ0 (z̄2 ),
+ o"→0 (1) + o r→0 (1).
"
y
Recall that the function y 7→ "χ(z̄2 , ψ0 (z̄2 ), " ) converges locally uniformly to y 7→ W (z̄2 , ψ0 (z̄2 ), y),
which is a Lipschitz continuous function, independent of y2 and such that W (z̄2 , ψ0 (z̄2 ), 0) = 0. Therefore, for " and r small enough, λϕ " (z ? ) + H̃"i (z ? , Dξ(z ? )) ≥ θ2 .
z?
?
If z" ∈ Γ , then, for some i ∈ {L, R}, H̃ +,i (z̄2 , Dξi (z ? ) − ψ0 (z2? )e2 + ψ0 (z̄2 )e2 , "2 ) ≥ E(z̄2 , ψ0 (z̄2 )). Since
the Hamiltonian H̃ +,i enjoys the same regularity properties as H̃ i , see Lemma 4.4.1, it is possible to use
?
+,i
the same arguments as in the case z" ∈ Ωi . For r and " small enough, λϕ " (z ? ) + H̃Γ ," (z ? , Dξ(z ? )) ≥ θ2 .
"
The claim that ϕ is a supersolution of (4.7.7) is proved.
Step 2
Let us prove that there exist some positive constants K r > 0 and "0 > 0 such that
ṽ " (z) + K r ≤ ϕ " (z),
∀z ∈ ∂ B(z̄, r), ∀" ∈ (0, "0 ).
(4.7.8)
Indeed, since ṽ−ϕ has a strict local maximum at z̄ and since ṽ(z̄) = ϕ(z̄), there exists a positive constant
K̃ r > 0 such that ṽ(z) + K̃ r ≤ ϕ(z) for any z ∈ ∂ B(z̄, r). Since ṽ = lim sup∗ ṽ" , there exists "˜0 > 0 such
"
that ṽ " (z) + 2r ≤ ϕ(z) for any 0 < " < "˜0 and z ∈ ∂ B(z̄, r). But z 7→ ϕ " (z) converges locally uniformly
to z 7→ ψ(z2 ) + W (z̄2 , ψ0 (z̄2 ), z) as " tends to 0. Hence, thanks to (4.5.14) in Corollary 4.5.2,
Š
€ R
L
ψ(z2 ) + W (z̄2 , ψ0 (z̄2 ), z) ≥ ψ(z2 ) + Π (z̄2 , ψ0 (z̄2 ))1z∈ΩR + Π (z̄2 , ψ0 (z̄2 ))1z∈Ω L z1 ,
K̃
and we get (4.7.8) for some constants K r > 0 and "0 > 0.
Step 3 From the previous steps and the local comparison principle in Theorem 4.2.1, we find that for
r and " small enough,
ṽ " (z) + K r ≤ ϕ " (z)
∀z ∈ B(z̄, r).
Taking z = z̄ and letting " → 0, we obtain
ṽ(z̄) + K r ≤ ψ(z̄2 ) = ϕ(z̄) = ṽ(z̄),
which cannot happen. The proof is completed.
Remark 4.7.1. For the proof of the supersolution inequality, the test-function ϕ should be chosen of the
form
b R (z̄2 , ψ0 (z̄2 ))z1 + 1Ω L Π
b L (z̄2 , ψ0 (z̄2 ))z1 ,
ϕ(z1 , z2 ) = ψ(z2 ) + 1ΩR Π
b i (z̄2 , ψ0 (z̄2 )) are defined in (4.4.27) and (4.4.29).
where ψ ∈ C 1 (R) and for i = L, R, Π
4.8
4.8.1
Appendix
Proof of Lemma 4.4.2
Subsolutions Let ϕ ∈ Rρ be a test-function and ȳ ∈ [−ρ, ρ] × R be such that u0 − ϕ has a strict local
maximum at ȳ. If ȳ ∈ (−ρ, 0) × R, (resp. ȳ ∈ (0, ρ) × R ) is standard to check that H̃ L ((0, z2 ), D( ȳ) +
p2 e2 , ȳ2 ) ≤ λ, (resp. H̃ R ((0, z2 ), D( ȳ) + p2 e2 , ȳ2 ) ≤ λ). We may focus on the case when ȳ = (0, ȳ2 ) ∈ Γ ,
because the cases ȳ1 = ±ρ can be treated with similar but simpler arguments. We wish to prove that
H̃ +,i ((0, z2 ), Dϕ i ( ȳ) + p2 e2 ȳ2 ) ≤ λ,
∀i = L, R.
(4.8.1)
4.8. APPENDIX
149
We may assume that for all η ∈ [0, 1], the function uη − ϕ is Lipschitz continuous in [−ρ, ρ] × R with
a Lipschitz constant L̄ independent of η. Fix i = L, R, we define the distance di to Ω̄i by
§
0
if y ∈ Ωi ,
di ( y) =
| y1 | otherwise.
Clearly, di ∈ Rρ . Take C = L̄ + 1. The function y 7→ u0 ( y) − ϕ( y) − C di ( y) has a strict local maximum
at ȳ. Thanks to the local uniform convergence of uη to u0 , there exists r ∈ (0, ρ) and a sequence of
points y η ∈ B( ȳ, r) such that
uη ( y) − ϕ( y) − C di ( y) ≤ uη ( y η ) − ϕ( y η ) − C di ( y η ),
for all y ∈ B( ȳ, r).
Up to the extraction of a subsequence, we may assume that y η → ȳ as η tends to 0. Note that y η ∈ Ω̄i .
η
Indeed, if it was not the case, then calling y η = (0, y2 ) ∈ B( ȳ, r),
η
uη ( y η ) − ϕ( y η ) − uη ( y η ) − ϕ( y η ) ≤ L̄| y1 | = L̄di ( y η ),
and uη ( y η )−ϕ( y η )− C di ( y η ) ≤ uη ( y η )−ϕ( y η )− di ( y η ) < uη ( y η )−ϕ( y η )− C di ( y η ), in contradiction
with the definition of y η .
Up to the extraction of subsequences, we can make out two cases:
η
Case 1: y η ∈ Γ . We obtain ηuη ( y η ) + maxi=L,R H̃ +,i ((0, z2 ), Dui ( y η ) + p2 e2 , y2 ) ≤ λη , and then
(4.8.1) by letting η → 0.
η
Case 2: y η ∈ Ωi . We obtain that ηuη ( y η ) + H̃ i ((0, z2 ), Dϕ( y η ) + p2 e2 , y2 ) ≤ λη , and then by letting
i
η → 0 that H̃ ((0, z2 ), Dϕ( ȳ) + p2 e2 , ȳ2 ) ≤ λ, from the continuity of the Hamiltonian H̃ i . Finally, since
H̃ i − H̃ +,i ≥ 0, which yields that H̃ +,i ((0, z2 ), Dϕ i ( ȳ) + p2 e2 ȳ2 ) ≤ λ.
Since the arguments above can be applied for i = L and i = R, we have obtained (4.8.1).
Supersolutions Let ϕ ∈ Rρ be a test-function and ȳ ∈ [−ρ, ρ] × R be such that u0 − ϕ has a strict
local minimum at ȳ. As above, we may focus on the case when ȳ ∈ Γ . We wish to prove that
max H̃ +,i ((0, z2 ), Dϕ i ( ȳ) + p2 e2 , ȳ2 ) ≥ λ.
(4.8.2)
i=L,R
Define
p̃0L
min
p̃0R
max
=
§
=
§
p∈R:
H̃ L ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 ) = H̃ +,L ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 )
ª
p∈R:
H̃ R ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 ) = H̃ +,R ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 )
ª
Recall that thanks to Lemma 4.3.1 and Remark 4.4.2,
§ L
Ẽ0 (z2 , ∂ y2 ϕ( ȳ) + p2 , ȳ2 )
H̃ +,L ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 ) =
H̃ 1 ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 )
§ 2
H̃ ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 )
H̃ +,R ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 ) =
Ẽ0R (z2 , ∂ y2 ϕ( ȳ) + p2 , ȳ2 )
,
if p ≤ p̃0L ,
if p ≥ p̃0L ,
if p ≤ p̃0R ,
if p ≥ p̃0R .
where Ẽ0i (z2 , p2 , y2 ) is defined in Remark 4.4.2.
We make out two cases:
Case 1: ∂ y1 ϕ L ( ȳ) ≥ p̃0L , ∂ y1 ϕ R ( ȳ) ≤ p̃0R and for each i ∈ {L, R}
H̃ i ((0, z2 ), Dϕ i ( ȳ) + p2 e2 , ȳ2 ) = max H̃ +, j ((0, z2 ), Dϕ j ( ȳ) + p2 e2 , ȳ2 ) .
j=L,R
(4.8.3)
In this case, we can use a standard stability argument: there exists a sequence y η of local minimum
points of uη − ϕ which converges to ȳ as η tends to 0. If, for a subsequence still called y η , y η ∈ Γ , then
150
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
η
ηuη ( y η ) + maxi=L,R H̃ +,i ((0, z2 ), Dϕ i ( y η ) + p2 e2 , y2 ) ≥ λη , because uη is a supersolution of (4.4.9),
and (4.8.2) is obtained by letting η → 0.
η
If for a subsequence, y η ∈ Ωi for some i ∈ {L, R}, then ηuη ( y η ) + H̃ i ((0, z2 ), Dϕ( y η ) + p2 e2 , y2 ) ≥ λη ,
and by letting η → 0, H̃ i ((0, z2 ), Dϕ i ( ȳ) + p2 e2 , ȳ2 ) ≥ λ. Finally, (4.8.2) is obtained thanks to (4.8.3).
Case 2: the assumptions of the case 1 are not satisfied. Thanks to the above identities for H̃ +,R ((0, z2 ), (∂ y2 ϕ( ȳ)+
p2 )e2 + pe1 , ȳ2 ) and H̃ +,L ((0, z2 ), (∂ y2 ϕ( ȳ) + p2 )e2 + pe1 , ȳ2 ), by modifying the slopes of ϕ in the normal
direction on each side of Γ , it is possible to construct a test-function ψ ∈ Rρ such that ψ( ȳ) = ϕ( ȳ),
∂ y2 ψ( ȳ) = ∂ y2 ϕ( ȳ), ∂ y1 ψ L ( ȳ) ≥ ∂ y1 ϕ L ( ȳ), ∂ y1 ψR ( ȳ) ≤ ∂ y1 ϕ R ( ȳ), and ψ satisfies (4.8.3) for each
i ∈ {L, R}. Thus, since ψ touches ϕ at ȳ from below, ȳ is still a strict local minimum point of u0 − ψ
and we conclude by applying the result proved in the first case.
4.8.2
Proof of Lemma 4.5.1
In the following lemma, we state some useful properties of the Hamiltonians H̃ i and H̃ −,i , in addition
to those already described in Lemma 4.4.1:
Lemma 4.8.1. For i = L, R, any z2 , y2 , y20 ∈ R and p ∈ R2 , the Hamiltonians H̃ i and H̃ ±,i , respectively
defined in (4.4.1) and (4.4.3), satisfy
| H̃
±,i
| H̃ i ((0, z2 ), p, y2 ) − H̃ i ((0, z2 ), p, y20 ) |≤ M̃ | y2 − y20 |
(4.8.4)
y20 |,
(4.8.5)
((0, z2 ), p, y2 ) − H̃
±,i
((0, z2 ), p,
y20 )
|≤ M̃ (C1 |p| + C2 ) | y2 −
where the constants M̃ , C1 and C2 can be computed from M` , M f , kg 0 k∞ , kg”k∞ and δ0 in Assumptions
[H0]-[H3].
Proof. The proof of (4.8.4) stems from the definition of H̃ i . Let us prove (4.8.5) for H̃ +,R , the other
cases being similar. Remark that H̃ +,R can be written
H̃ +,R ((0, z2 ), p, y2 ) =
a∈Ai
s.t.
max
(− f˜z2 ( y2 , a) · p − ˜`z2 ( y2 , a)),
f˜z2 ( y2 ,a)·e1 ≥0
with f˜z2 ( y2 , a) = J̃( y2 ) f R ((0, z2 ), a) and ˜`z2 ( y2 , a) = `R ((0, z2 ), a). It is a simple matter to check that
f˜z2 and ˜`z2 satisfy the Assumptions [H0]-[H3] with different constants L f˜ , M˜` , δ̃0 and the modulus
ω̃` (·) = 0. Hence, the proof of [Oud14, Lemma 3.6] can be repeated to get (4.8.5).
Let us now prove Lemma 4.5.1. We follow the lines of [BCD97], Theorem IV.5.8 and argue by
contradiction, assuming that
m=
sup
(u( y) − v( y)) > 0.
y∈[ρ1 ,ρ2 ]×R
Then, for any η ∈ (0, m), there exists ξ ∈ [ρ1 , ρ2 ] × R such that
u(ξ) − v(ξ) ≥ m − η > 0.
(4.8.6)
We fix a such pair (η, ξ). From the inequalities on the boundary { y1 = ρ1 }, we see that ξ1 > ρ1 . Let
us proceed by doubling the variables and introduce the auxiliary function
2
x − y
ψ" (x, y) = u(x) − v( y) − − γ(ξ) − | y − ξ |2 ,
"
where " is a positive parameter and γ : [ρ1 , ρ2 ] × R → R is given by

3ρ +ρ
if ζ1 ∈ [ρ1 , 14 2 ],
 (1, 0)
−4ζ +2(ρ +ρ )
3ρ +ρ ρ +3ρ
γ(ζ1 , ζ2 ) =
( 1ρ2 −ρ11 2 , 0) if ζ1 ∈ [ 14 2 , 1 4 2 ],

ρ +3ρ
(−1, 0)
if ζ1 ∈ [ 1 4 2 , ρ2 ],
Take c = 41 min(ρ2 − ρ1 , 1). It is easy to check that for all ζ ∈ [ρ1 , ρ2 ] × R and t ∈ (0, c], B(ζ +
tγ(ζ), c t)) ⊂ (ρ1 , ρ2 ) × R. Let (x " , y " ) ∈ [ρ1 , ρ2 ] × R be a point where ψ" reaches its maximum.
We claim that for η small enough and for a subsequence, ρ1 < x 1" < ρ2 and ρ1 < y1" ≤ ρ2 . Indeed,
ψ" (ξ + "γ(ξ), ξ) = u(ξ + "γ(ξ)) − v(ξ) ≥ m − η + u(ξ + "γ(ξ)) − u(ξ).
4.8. APPENDIX
Since lim"→0 u(ξ + "γ(ξ)) = u(ξ), for " small enough, ψ" (ξ + "γ(ξ), ξ) ≥ m −
|
151
3η
2 .
Therefore,
x" − y"
3η
− γ(ξ) |2 + | y " − ξ |2 ≤
− m + u(x " ) − v( y " ).
"
2
(4.8.7)
Since the right side of (4.8.7) is bounded, we may assume up to the extraction of a subsequence that
(x " , y " ) → (x ? , x ? ) as " tends to 0. From the continuity of u and v, lim"→0 u(x " ) − v( y " ) ≤ u(x ? ) −
v(x ? ) ≤ m. For the subsequence, (4.8.7) implies that
|
x" − y"
3η
− γ(ξ) |2 + | y " − ξ |2 ≤
+ o" (1),
"
2
(4.8.8)
Since ξ1 > ρ1 ,(4.8.8) implies that y1" > ρ1 for η and " small enough. To prove that ρ1 < x 1" < ρ2 , we
observe that there exists θ > 0 such that for any x, y ∈ [ρ1 , ρ2 ] × R, | γ(x) − γ( y) |≤ 2c if |x − y| ≤ θ .
2
For η and " small enough so that the right hand side of (4.8.8) is smaller than min{θ 2 , c4 }, we check
that | y " − ξ| ≤ θ and consequently |γ( y " ) − γ(ξ)| ≤ 2c . Then (4.8.8) yields that
| x " − y " − "γ( y " ) |≤ "c,
and therefore that x " ∈ (ρ1 , ρ2 ) × R and the claim is proved.
Let us choose the test-functions
ϕ1 (x)
=
ϕ2 ( y)
=
x − y"
− γ(ξ) |2 ,
"
x" − y
− γ(ξ) |2 − | y − ξ |2 .
−|
"
|
Since u is a subsolution of (4.5.2) and u − ϕ1 has local maximum at x " ,


‹
‹
2 x" − y"
H̃ R (0, z2 ),
− γ(ξ) + p2 e2 , x 2" ≤ λ − "0 .
"
"
For the supersolution inequality, we make out two cases:
• If y " ∈ (ρ1 , ρ2 ) × R, then

‹

‹
2 x" − y"
R
"
"
H̃ (0, z2 ),
− γ(ξ) − 2 ( y − ξ) + p2 e2 , y2 ≥ λ,
"
"
(4.8.9)
(4.8.10)
because v is a supersolution of (4.5.1) and v − ϕ2 has a local minimum at y " . Subtracting (4.8.9) from
(4.8.10), we get


‹
‹


‹
‹
2 x" − y"
2 x" − y"
R
R
"
"
"
H̃ (0, z2 ),
− γ(ξ) − 2 ( y − ξ) + p2 e2 , y2 − H̃ (0, z2 ),
− γ(ξ) + p2 e2 , x 2
"
"
"
"
≥ "0 .
(4.8.11)
Thus, thanks to Lemmas 4.8.1 and 4.4.1, "0 ≤ M̃ |x 2" − y2" | + 2M f | y " − ξ|, in contradiction with (4.8.8)
for η and " small enough.
• If y " ∈ {ρ2 } × R, then


‹
‹
2 x" − y"
H̃ −,R (0, z2 ),
− γ(ξ) − 2 ( y " − ξ) + p2 e2 , y2" ≥ λ
(4.8.12)
"
"
because v is a supersolution of (4.5.1) and v −ϕ2 has a local minimum at y " . Since H̃ −,R ≤ H̃ R , (4.8.12)
implies (4.8.10) and we may conclude as in the first case.
4.8.3
Proof of Theorem 4.7.1
The proof of Theorem 4.7.1 follows the lines of [IM14a, Theorem 2.6] and relies on the following two
technical lemmas, which can be proved by adapting the proofs of Lemma 2.8 and Lemma 2.9 in [IM14a],
with very slight changes.
152
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
Lemma 4.8.2. Let u : R2 → R be a subsolution of (4.1.21) and φ ∈ R touching u from above at
z̄ = (0, z̄2 ) ∈ Γ . For each i ∈ {L, R}, the real number p̄i :
p̄i = inf p ∈ R : ∃r > 0 s.t. φ(z) + σ i pz1 ≥ u(z), ∀z = (z1 , z2 ) ∈ [0, r) × (z̄2 − r, z̄2 + r) ,
where σ i is given by (4.1.5), is nonpositive. Moreover,
λu(0, z̄2 ) + H i ((0, z̄2 ), Dφ i (0, z̄2 ) + σ i p̄i e1 ) ≤ 0.
(4.8.13)
Lemma 4.8.3. Let w : R2 → R be a supersolution of (4.1.21) and φ ∈ R touching w from below at
z̄ = (0, z̄2 ) ∈ Γ . For each i ∈ {L, R}. For each i ∈ {L, R}, the real number p̃i :
p̃i = sup p ∈ R : ∃r > 0 s.t. φ(z) + σ i pz1 ≤ w(z), ∀z = (z1 , z2 ) ∈ [0, r) × (z̄2 − r, z̄2 + r) ,
is nonnegative. Moreover
λw(0, z̄2 ) + H i ((0, z̄2 ), Dφ i (0, z̄2 ) + σ i p̃i e1 ) ≥ 0.
(4.8.14)
We are now ready to prove Theorem 4.7.1.
Subsolutions Let Π be a map as in Definition 4.7.1. Suppose that a subsolution u of (4.1.21) satisfies
(4.7.3) for all z ∈ Γ and all test-functions in R Π (z) touching u from above at z.
Let φ ∈ R be such that u − φ has a strict local maximum at z̄ ∈ Γ and that u(z̄) = φ(z̄). We wish to
prove that
λu(z̄) + max E(z̄2 , ∂z2 φ(z̄)), HΓ (z, Dφ L (z̄), Dφ R (z̄)) ≤ 0.
(4.8.15)
From Lemma 4.8.2, for each i ∈ {L, R}, there exists p̄i ≤ 0 such that
λu(z̄) + H i (z̄, Dφ i (z̄) + σ i p̄i e1 ) ≤ 0.
(4.8.16)
From the monotonicity properties of the Hamiltonians H +,i stated in Lemma 4.3.1,
HΓ (z, Dφ L (z̄), Dφ R (z̄)) ≤ HΓ (z, Dφ L (z̄) − p̄ L e1 , Dφ R (z̄) + p̄R e1 )
≤ max H L (z, Dφ L (z̄) − p̄ L e1 ), H R (z, Dφ R (z̄) + p̄R e1 ) .
Hence, from (4.8.16),
λu(z̄) + HΓ (z̄, Dφ L (z̄), Dφ R (z̄)) ≤ 0.
Therefore, in order to prove (4.8.15), we are left with checking that
λu(z̄) + E(z̄2 , ∂z2 φ(z̄)) ≤ 0.
(4.8.17)
Recall that from Proposition 4.4.1, E(·, ·) ≥ E0 (·, ·). If E(z̄2 , ∂z2 φ(z̄)) = E0 (z̄2 , ∂z2 φ(z̄)), then (4.8.17) is
a direct consequence of (4.8.16). Let us consider the case when E(z̄2 , ∂z2 φ(z̄)) > E0 (z̄2 , ∂z2 φ(z̄)) and
assume by contradiction that
λu(z̄) + E(z̄2 , ∂z2 φ(z̄)) > 0.
(4.8.18)
Then, from (4.8.16) , for any i ∈ {L, R},
H −,i (z̄, Dφ i (z̄) + σ i p̄i e1 ) ≤ −λu(z̄) < E(z̄2 , ∂z2 φ(z̄)).
From this and the monotonicity properties of the functions p ∈ R 7→ H −,i (z, ∂z2 φ(z̄)e2 + pe1 ), we deduce
that
Π L (z̄2 , ∂z2 φ(z̄)) < ∂z1 φ L (z̄) − p̄ L , and ΠR (z̄2 , ∂z2 φ(z̄)) > ∂z1 φ R (z̄) + p̄R .
Thus, the modified test-function ϕ ∈ R Π (z̄) defined by
ϕ(z) = φ(0, z2 ) + 1Ω L (z)Π L (z̄2 , ∂z2 φ(z̄))z1 + 1ΩR (z)ΠR (z̄2 , ∂z2 φ(z̄))z1
is such that u − ϕ has a local maximum at z̄, and therefore
λu(z̄) + max E(z̄2 , ∂z2 ϕ(z̄)), HΓ (z̄, Dϕ L (z̄), Dϕ R (z̄)) ≤ 0,
which contradicts (4.8.18).
4.8. APPENDIX
153
Supersolutions Suppose that a supersolution u of (4.1.21) satisfies (4.7.3) for all z ∈ Γ and all testfunctions in R Π (z) touching u from below at z.
Let φ ∈ R be such that u − φ has a strict local maximum at z̄ ∈ Γ with u(z̄) = φ(z̄). We wish to prove
that
λu(z̄) + max E(z̄2 , ∂z2 φ(z̄)), HΓ (z, Dφ L (z̄), Dφ R (z̄)) ≥ 0.
(4.8.19)
From Lemma 4.8.3, for each i ∈ {L, R} there exists p̃i ≥ 0 such that
λu(z̄) + H i (z̄, Dφ i (z̄) + σ i p̃i e1 ) ≥ 0,
(4.8.20)
and using the monotonicity properties of the Hamiltonians H +,i , see Lemma 4.3.1,
HΓ (z̄, Dφ L (z̄), Dφ R (z̄)) ≥ HΓ (z̄, Dφ L (z̄) − p̃ L e1 , Dφ R (z̄) + p̃R e1 ).
(4.8.21)
If for some i ∈ {L, R}, H +,i (z̄, Dφ i (z̄) + σ i p̃i e1 ) = H i (z̄, Dφ i (z̄) + σ i p̃i e1 ), then (4.8.19) follows readily
from (4.8.20) and (4.8.21) so we may now suppose that
H i (z̄, Dφ i (z̄) + σ i p̃i e1 ) = H −,i (z̄, Dφ i (z̄) + σ i p̃i e1 ),
i = L, R.
(4.8.22)
Assume by contradiction that λu(z̄) + max E(z̄2 , ∂z2 φ(z̄)), HΓ (z̄, Dφ L (z̄), Dφ R (z̄)) < 0. Thus, from
(4.8.20) and (4.8.22),
E(z̄2 , ∂z2 φ(z̄)) < −λu(z̄) ≤ H −,i (z̄, Dφ i (z̄) + σ i p̃i e1 ),
i = L, R.
(4.8.23)
From (4.8.22), (4.8.23) and the monotonicity properties of the functions p ∈ R 7→ H −,i (z, ∂z2 φ(z̄)e2 +
pe1 ), we deduce that
Π L (z̄2 , ∂z2 φ(z̄)) > ∂z1 φ L (z̄) − p̃ L ,
and ΠR (z̄2 , ∂z2 φ(z̄)) < ∂z1 φ R (z̄) + p̃R .
Since the modified test-function ϕ ∈ R Π (z̄),
ϕ(z) = φ(0, z2 ) + 1Ω L Π L (z̄2 , ∂z2 φ(z̄))z1 + 1ΩR ΠR (z̄2 , ∂z2 φ(z̄))z1 ,
is such that u − ϕ has a local minimum at z̄, we get
λu(z̄) + max E(z̄2 , ∂z2 φ(z̄)), HΓ (z, Dϕ L (z̄), Dϕ R (z̄)) ≥ 0,
which is the desired contradiction.
154
CHAPITRE 4. AN OSCILLATING INTERFACE PROBLEM
BIBLIOGRAPHY
[AB02]
O. Alvarez and M. Bardi. Viscosity solutions methods for singular perturbations in deterministic and stochastic control. SIAM J. Control Optim., 40(4):1159–1188 (electronic),
2001/02.
[AB03]
O. Alvarez and M. Bardi. Singular perturbations of nonlinear degenerate parabolic PDEs:
a general convergence result. Arch. Ration. Mech. Anal., 170(1):17–61, 2003.
[ABM08]
O. Alvarez, M. Bardi, and C. Marchi. Multiscale singular perturbations and homogenization
of optimal control problems. In Geometric control and nonsmooth analysis, volume 76 of
Ser. Adv. Math. Appl. Sci., pages 1–27. World Sci. Publ., Hackensack, NJ, 2008.
[ACCT13]
Y. Achdou, F. Camilli, A. Cutrì, and N. Tchou. Hamilton-Jacobi equations constrained on
networks. NoDEA Nonlinear Differential Equations Appl., 20(3):413–445, 2013.
[AF14]
L. Ambrosio and J. Feng. On a class of first order Hamilton-Jacobi equations in metric
spaces. J. Differential Equations, 256(7):2194–2245, 2014.
[AG00]
Z. Artstein and V. Gaitsgory. The value function of singularly perturbed control systems.
Appl. Math. Optim., 41(3):425–445, 2000.
[AGS08]
L. Ambrosio, N. Gigli, and G. Savaré. Gradient flows in metric spaces and in the space
of probability measures. Lectures in Mathematics ETH Zürich. Birkhäuser Verlag, Basel,
second edition, 2008.
[AKR11]
B. Andreianov, K. H. Karlsen, and N. H. Risebro. A theory of L 1 -dissipative solvers for scalar
conservation laws with discontinuous flux. Arch. Ration. Mech. Anal., 201(1):27–86, 2011.
[AOT15a]
Y. Achdou, S. Oudet, and N. Tchou. Effective transmission conditions for Hamilton-Jacobi
equations defined on two domains separated by an oscillatory interface. arXiv preprint
arXiv:1506.02998, 2015.
[AOT15b]
Y. Achdou, S. Oudet, and N. Tchou. Hamilton-Jacobi equations for optimal control on
junctions and networks. published online http://dx.doi.org/10.1051/cocv/2014054, May
2015.
[AT15]
Y. Achdou and N. Tchou. Hamilton-Jacobi equations on networks as limits of singularly
perturbed problems in optimal control: dimension reduction. Comm. Partial Differential
Equations, 40(4):652–693, 2015.
[Bar94]
G. Barles. Solutions de viscosité des équations de Hamilton-Jacobi, volume 17 of Mathématiques & Applications (Berlin) [Mathematics & Applications]. Springer-Verlag, Paris, 1994.
[Bar13]
G. Barles. An introduction to the theory of viscosity solutions for first-order HamiltonJacobi equations and applications. In Hamilton-Jacobi equations: approximations, numerical analysis and applications, volume 2074 of Lecture Notes in Math., pages 49–109.
Springer, Heidelberg, 2013.
[BBC13]
G. Barles, A. Briani, and E. Chasseigne. A Bellman approach for two-domains optimal
control problems in RN . ESAIM Control Optim. Calc. Var., 19(3):710–739, 2013.
[BBC14]
G. Barles, A. Briani, and E. Chasseigne. A Bellman approach for regional optimal control
problems in RN . SIAM J. Control Optim., 52(3):1712–1744, 2014.
[BBCT14]
G. Barles, A. Briani, E Chasseigne, and N. Tchou. Homogenization Results for a Deterministic Multi-domains Periodic Control Problem. May 2014.
155
156
BIBLIOGRAPHY
[BC14]
G. Barles and E. Chasseigne. (Almost) Everything You Always Wanted to Know About
Deterministic Control Problems in Stratified Domains. arXiv preprint arXiv:1412.7556,
2014.
[BCD97]
M. Bardi and I. Capuzzo-Dolcetta. Optimal control and viscosity solutions of HamiltonJacobi-Bellman equations. Systems & Control: Foundations & Applications. Birkhäuser
Boston, Inc., Boston, MA, 1997. With appendices by Maurizio Falcone and Pierpaolo Soravia.
[BCF51]
W. K. Burton, N. Cabrera, and F. C. Frank. The growth of crystals and the equilibrium
structure of their surfaces. Philos. Trans. Roy. Soc. London. Ser. A., 243:299–358, 1951.
[Ben88]
A. Bensoussan. Perturbation methods in optimal control. Wiley/Gauthier-Villars Series in
Modern Applied Mathematics. John Wiley & Sons Ltd., Chichester, 1988. Translated from
the French by C. Tomson.
[BH07]
A. Bressan and Y. Hong. Optimal control problems on stratified domains. Netw. Heterog.
Media, 2(2):313–331, 2007.
[BHN+ 95] M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y. Sugiyama. Dynamical model of
traffic congestion and numerical simulation. Physical Review E, 51(2):1035, 1995.
[BJ02]
G. Barles and E. R. Jakobsen. On the convergence rate of approximation schemes for
Hamilton-Jacobi-Bellman equations. M2AN Math. Model. Numer. Anal., 36(1):33–54, 2002.
[Bla97]
A.-P. Blanc. Deterministic exit time control problems with discontinuous exit costs. SIAM
J. Control Optim., 35(2):399–434, 1997.
[Bla01]
A.-P. Blanc. Comparison principle for the Cauchy problem for Hamilton-Jacobi equations
with discontinuous data. Nonlinear Anal., 45(8, Ser. A: Theory Methods):1015–1037,
2001.
[BP90]
G. Barles and B. Perthame. Comparison principle for Dirichlet-type Hamilton-Jacobi equations and singular perturbations of degenerated elliptic equations. Appl. Math. Optim.,
21(1):21–44, 1990.
[BS91]
G. Barles and P. E. Souganidis. Convergence of approximation schemes for fully nonlinear
second order equations. Asymptotic Anal., 4(3):271–283, 1991.
[BW13]
R. C. Barnard and P. R. Wolenski. Flow invariance on stratified domains. Set-Valued Var.
Anal., 21(2):377–403, 2013.
[CCM14]
Fabio Camilli, Raffaela Capitanelli, and Claudio Marchi. Eikonal equations on the sierpinski
gasket. arXiv preprint arXiv:1404.3692, 2014.
[CDL90]
I. Capuzzo-Dolcetta and P.-L. Lions. Hamilton-Jacobi equations with state constraints.
Trans. Amer. Math. Soc., 318(2):643–683, 1990.
[CIL92]
M. G. Crandall, H. Ishii, and P-L. Lions. User’s guide to viscosity solutions of second order
partial differential equations. Bull. Amer. Math. Soc. (N.S.), 27(1):1–67, 1992.
[CL81]
M. G. Crandall and P.-L. Lions. Condition d’unicité pour les solutions généralisées des équations de Hamilton-Jacobi du premier ordre. C. R. Acad. Sci. Paris Sér. I Math., 292(3):183–
186, 1981.
[CL83]
M. G. Crandall and P.-L. Lions. Viscosity solutions of Hamilton-Jacobi equations. Trans.
Amer. Math. Soc., 277(1):1–42, 1983.
[CL84]
M. G. Crandall and P.-L. Lions. Two approximations of solutions of Hamilton-Jacobi equations. Math. Comp., 43(167):1–19, 1984.
[CM13]
F. Camilli and C. Marchi. A comparison among various notions of viscosity solution for
Hamilton-Jacobi equations on networks. J. Math. Anal. Appl., 407(1):112–118, 2013.
BIBLIOGRAPHY
157
[CR07]
G. M. Coclite and N. H. Risebro. Viscosity solutions of Hamilton-Jacobi equations with
discontinuous coefficients. J. Hyperbolic Differ. Equ., 4(4):771–795, 2007.
[CS03]
F. Camilli and A. Siconolfi. Hamilton-Jacobi equations with measurable dependence on the
state variable. Adv. Differential Equations, 8(6):733–768, 2003.
[CS05]
F. Camilli and A. Siconolfi. Time-dependent measurable Hamilton-Jacobi equations. Comm.
Partial Differential Equations, 30(4-6):813–847, 2005.
[CSM13]
F. Camilli, D. Schieborn, and C. Marchi. Eikonal equations on ramified spaces. Interfaces
Free Bound., 15(1):121–140, 2013.
[Dup92]
P. Dupuis. A numerical method for a calculus of variations problem with discontinuous
integrand. In Applied stochastic analysis (New Brunswick, NJ, 1991), volume 177 of Lecture
Notes in Control and Inform. Sci., pages 90–107. Springer, Berlin, 1992.
[EKFNS08] K.-J. Engel, M. Kramar Fijavž, R. Nagel, and E. Sikolya. Vertex control of flows in networks.
Netw. Heterog. Media, 3(4):709–722, 2008.
[Eva89]
L. C. Evans. The perturbed test function method for viscosity solutions of nonlinear PDE.
Proc. Roy. Soc. Edinburgh Sect. A, 111(3-4):359–375, 1989.
[FP00a]
H. Frankowska and S. Plaskacz. Hamilton-Jacobi equations for infinite horizon control
problems with state constraints. In Calculus of variations and optimal control (Haifa, 1998),
volume 411 of Res. Notes Math., pages 97–116. Chapman & Hall/CRC, Boca Raton, FL,
2000.
[FP00b]
H. Frankowska and S. Plaskacz. Semicontinuous solutions of Hamilton-Jacobi-Bellman
equations with degenerate state constraints. J. Math. Anal. Appl., 251(2):818–838, 2000.
[FR14]
N. Forcadel and Z. Rao. Singular perturbation of optimal control problems on multidomains. SIAM J. Control Optim., 52(5):2917–2943, 2014.
[FS14a]
N. Forcadel and W. Salazar. A junction condition by specified homogenization of a discrete
model with a local perturbation and application to traffic flow. December 2014.
[FS14b]
N. Forcadel and W. Salazar. Homogenization of second order discrete model and application
to traffic flow. preprint, 2014.
[GGR11]
Y. Giga, P. Górka, and P. Rybka. A comparison principle for Hamilton-Jacobi equations with
discontinuous Hamiltonians. Proc. Amer. Math. Soc., 139(5):1777–1785, 2011.
[GH13]
Y. Giga and N. Hamamuki. Hamilton-Jacobi equations with discontinuous source terms.
Comm. Partial Differential Equations, 38(2):199–243, 2013.
[GHN14]
Y. Giga, N. Hamamuki, and A. Nakayasu. Eikonal equations in metric spaces. Trans. Amer.
Math. Soc., (1):49––66, 2014.
[GIM14]
G. Galise, C. Imbert, and R. Monneau. A junction condition by specified homogenization
and application to traffic lights. 41 pages, September 2014.
[GL99]
V. Gaitsgory and A. Leizarowitz. Limit occupational measures set for a control system and
averaging of singularly perturbed control systems. J. Math. Anal. Appl., 233(2):461–475,
1999.
[GNPT07] M. Garavello, R. Natalini, B. Piccoli, and A. Terracina. Conservation laws with discontinuous flux. Netw. Heterog. Media, 2(1):159–179, 2007.
[GP06]
M. Garavello and B. Piccoli. Traffic flow on networks, volume 1 of AIMS Series on Applied
Mathematics. American Institute of Mathematical Sciences (AIMS), Springfield, MO, 2006.
Conservation laws models.
158
BIBLIOGRAPHY
[GS04]
M. Garavello and P. Soravia. Optimality principles and uniqueness for Bellman equations
of unbounded control problems with discontinuous running cost. NoDEA Nonlinear Differential Equations Appl., 11(3):271–298, 2004.
[GS06]
M. Garavello and P. Soravia. Representation formulas for solutions of the HJI equations with
discontinuous coefficients and existence of value in differential games. J. Optim. Theory
Appl., 130(2):209–229, 2006.
[HZ15]
C. Hermosilla and H. Zidani. Infinite horizon problems on stratifiable state-constraints sets.
J. Differential Equations, 258(4):1430–1460, 2015.
[IK96]
H. Ishii and S. Koike. A new formulation of state constraint problems for first-order PDEs.
SIAM J. Control Optim., 34(2):554–571, 1996.
[IM14a]
C. Imbert and R. Monneau. Flux-limited solutions for quasi-convex Hamilton-Jacobi equations on networks. https://hal.archives-ouvertes.fr/hal-00832545, October 2014.
[IM14b]
C. Imbert and R. Monneau. Quasi-convex Hamilton-Jacobi equations posed on junctions:
the multi-dimensional case. 18 pages. Preliminary version., October 2014.
[IMZ13]
C. Imbert, R. Monneau, and H. Zidani. A Hamilton-Jacobi approach to junction problems
and application to traffic flows. ESAIM Control Optim. Calc. Var., 19(1):129–166, 2013.
[Ish85]
H. Ishii. Hamilton-Jacobi equations with discontinuous Hamiltonians on arbitrary open
sets. Bull. Fac. Sci. Engrg. Chuo Univ., 28:33–77, 1985.
[Ish87a]
H. Ishii. Perron’s method for Hamilton-Jacobi equations. Duke Math. J., 55(2):369–384,
1987.
[Ish87b]
H. Ishii. A simple, direct proof of uniqueness for solutions of the Hamilton-Jacobi equations
of eikonal type. Proc. Amer. Math. Soc., 100(2):247–251, 1987.
[Ish89]
H. Ishii. A boundary value problem of the Dirichlet type for Hamilton-Jacobi equations.
Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 16(1):105–135, 1989.
[Ish13]
H. Ishii. A short introduction to viscosity solutions and the large time behavior of solutions
of Hamilton-Jacobi equations. In Hamilton-Jacobi equations: approximations, numerical
analysis and applications, volume 2074 of Lecture Notes in Math., pages 111–249. Springer,
Heidelberg, 2013.
[Jen88]
R. Jensen. The maximum principle for viscosity solutions of fully nonlinear second order
partial differential equations. Arch. Rational Mech. Anal., 101(1):1–27, 1988.
[JLS88]
R. Jensen, P.-L. Lions, and P. E. Souganidis. A uniqueness result for viscosity solutions of second order fully nonlinear partial differential equations. Proc. Amer. Math. Soc., 102(4):975–
978, 1988.
[Lio82]
P.-L. Lions. Generalized solutions of Hamilton-Jacobi equations, volume 69 of Research Notes
in Mathematics. Pitman (Advanced Publishing Program), Boston, Mass.-London, 1982.
[Lio14]
P-L. Lions.
Cours du Collège de France.
http://www.college-defrance.fr/default/EN/all/equ− der/, january and february 2014.
[LPV]
P.-L. Lions, G. Papanicolaou, and S.R.S Varadhan. Homogenization of Hamilton-Jacobi
equations. unpublished.
[MW67]
E. J. McShane and R. B. Warfield, Jr. On Filippov’s implicit functions lemma. Proc. Amer.
Math. Soc., 18:41–47, 1967.
[Nak14]
A. Nakayasu. On metric viscosity solutions for Hamilton-Jacobi equations of evolution type.
arXiv preprint arXiv:1407.7634, 2014.
BIBLIOGRAPHY
159
[Oud14]
S. Oudet. Hamilton-Jacobi equations for optimal control on multidimensional junctions.
ArXiv e-prints, 2014.
[RSZ14]
Z. Rao, A. Siconolfi, and H. Zidani. Transmission conditions on interfaces for HamiltonJacobi-Bellman equations. J. Differential Equations, 257(11):3978–4014, 2014.
[RZ13]
Z. Rao and H. Zidani. Hamilton-jacobi-bellman equations on multi-domains. In Control
and Optimization with PDE Constraints, pages 93–116. Springer, 2013.
[SC13]
D. Schieborn and F. Camilli. Viscosity solutions of Eikonal equations on topological networks. Calc. Var. Partial Differential Equations, 46(3-4):671–686, 2013.
[Sch06]
D. Schieborn. Viscosity Solutions of Hamilton-Jacobi Equations of Eikonal Type on Ramified
Spaces. PhD thesis, Univ. Tübingen, 2006.
[Son86a]
H. M. Soner. Optimal control with state-space constraint. I. SIAM J. Control Optim.,
24(3):552–561, 1986.
[Son86b]
H. M. Soner. Optimal control with state-space constraint. II. SIAM J. Control Optim.,
24(6):1110–1122, 1986.
[Sor02]
P. Soravia. Boundary value problems for Hamilton-Jacobi equations with discontinuous
Lagrangian. Indiana Univ. Math. J., 51(2):451–477, 2002.
[Sor06]
P. Soravia. Degenerate eikonal equations with discontinuous refraction index. ESAIM Control Optim. Calc. Var., 12(2):216–230 (electronic), 2006.
[SZN+ 10]
G. Sazaki, S. Zepeda, S. Nakatsubo, E. Yokoyama, and Y. Furukawa. Elementary steps at
the surface of ice crystals visualized by advanced optical microscopy. Proceedings of the
National Academy of Sciences, 107(46):19702–19707, 2010.
[Val69]
M. Valadier. Sous-différentiels d’une borne supérieure et d’une somme continue de fonctions convexes. C. R. Acad. Sci. Paris Sér. A-B, 268:A39–A42, 1969.