Asymptotic distribution of a time-inhomogeneous

kinetic model
Emeline Luirard

To cite this version:

Emeline Luirard. Asymptotic distribution of a time-inhomogeneous kinetic model. Analysis of PDEs
[[Link]]. Université de Rennes, 2022. English. �NNT : 2022REN1S041�. �tel-03903575�

HAL Id: tel-03903575

[Link]
Submitted on 16 Dec 2022

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
T HÈSE DE DOCTORAT DE

L’UNIVERSITÉ DE RENNES 1

É COLE D OCTORALE N O 601

Mathématiques et Sciences et Technologies
de l’Information et de la Communication
Spécialité : Mathématiques et leurs Interactions

Par
Emeline L UIRARD
Loi limite de modèles cinétiques inhomogènes en temps

Thèse présentée et soutenue à Rennes, le 27 Juin 2022

Unité de recherche : IRMAR (UMR CNRS 6625)

Rapporteurs avant soutenance :

Patrick C ATTIAUX Professeur, Université Paul Sabatier
Nicolas P RIVAULT Professeur, Nanyang Technological University

Composition du Jury :
Présidente : Fabienne C ASTELL Professeure, Université Aix-Marseille
Examinateurs : Vlad B ALLY Professeur, Université Gustave Eiffel
Nils B ERGLUND Professeur, Université d’Orléans
Fabienne C ASTELL Professeure, Université Aix-Marseille
Camille TARDIF Maître de conférence, Sorbonne Université
Dir. de thèse : Mihai G RADINARU Professeur, Université de Rennes 1
“Je suis moi et les décisions d’une montagne pour imposante qu’elle soit n’y peuvent
rien ! Je suis moi et j’ai l’intention de poursuivre mon chemin où qu’il me mène !”

— Ellana, Le Pacte des Marchombres de Pierre Bottero

 R EMERCIEMENTS

C’est une nouvelle page qui se tourne, et je suis fière d’être arrivée jusqu’au bout de
cette aventure ! J’aimerais remercier toutes les personnes qui m’ont permis de réussir, de
près ou de loin. Je vais faire de mon mieux pour n’oublier personne.

J’exprime tout d’abord ma reconnaissance à Mihai Gradinaru, mon directeur de

thèse, pour son encadrement tout au long de mon doctorat et pour le temps qu’il m’a
consacré. Merci pour ta gentillesse et ta grande disponibilité. Merci également de m’avoir
fait découvrir le monde de la recherche.
Je remercie Patrick Cattiaux et Nicolas Privault de m’avoir fait l’honneur de
rapporter mon mémoire de thèse. Merci pour vos relectures précieuses et soignées. Merci
également à Vlad Bally, Nils Berglund, Fabienne Castell et Camille Tardif d’avoir
accepté d’être dans le jury.
Je tiens également à remercier tout particulièrement Jürgen Angst pour nos échanges,
ses nombreux conseils et son soutien. Merci à Francois Bolley et Karine Beauchard
d’avoir veillé de loin ces derniers mois.

Bien sûr je ne serai pas là où j’en suis aujourd’hui si je ne prenais pas plaisir à
faire des mathématiques ! Alors merci également à [Link] les (autres) [Link]
de mathématiques qui m’ont inspirée tout au long de ma scolarité. Pour n’en citer
que quelques-uns, merci à Mr Moine et ses carrés magiques, Mme Martinent, Mr
Gouttenoire, Judicaël Courant pour m’avoir appris la rigueur mathématique, Bruno
Arsac (pour tes encouragements tout au long de mon parcours), Thibault Deheuvels,
Jérémy Leborgne, Jean-Christophe Breton, Matthieu Romagny, Félix Ulmer, Hélène
Guerrin, Benjamin Boutin, Dominique Cerveau.
J’ai effectué ma thèse dans le cadre agréable de la tour des maths. J’en remercie
tous les membres, [Link], [Link], bibliothécaires, secrétaires. Merci tout
particulièrement à Adeline, Bernard, Barbara, Rozenn, Nathalie, Hélène, Marie-Aude,
Éric, Olivier, Dominique.
J’ai également eu la chance de pouvoir enseigner les mathématiques en licence de bi-

ii
ologie. Ces moments d’oxygène ont été essentiels pendant ces trois années. J’ai toujours
pris plaisir à enseigner et je remercie mes é[Link] d’avoir entretenu cette bonne am-
biance. Merci également à Jean-Christophe de m’avoir fait confiance pour préparer les
é[Link] de la prépa agreg à l’option Proba-Stats.

Bien sûr, je désire remercier les personnes avec qui j’ai partagé le bureau 334 pendant
ces trois années, Markus, Thibault, Marie, Louis et Marc. Merci pour les bons moments
partagés, les balades, les délicieux gâteaux, les discussions sur la recherche, l’enseignement
ou encore la danse, les conseils en pâtisserie, et surtout vos encouragements et votre
soutien. La crêperie du bureau 334 aura embaumé le couloir quelques mois ! Je pense
aussi à nos décomptes en dessin avant les vacances et aux beaux dessins de géométrie qui
fleurissent sur le tableau depuis quelques mois !
Je pense également à Josselin. Tes promenades parmi les étages t’ont souvent mené au
bureau 334. Merci pour tous tes bons conseils et tes origamis qui décorent mon bureau.
Merci Fabrice pour ta bonne humeur et tes conseils d’ancien. Je ne peux oublier le bureau
232, terre d’accueil. Merci à François, Pierre, Rémi et Thomas pour les discussions à refaire
le monde, les bons moments. Merci Thomas pour tes relectures précieuses et toutes nos
discussions. Faire des maths à plusieurs, c’est tellement mieux ! Merci également à Laura
et Florian pour avoir partagé toutes les bonnes infos et pour vos conseils avisés.
Je tiens à remercier Alice d’être là et de m’avoir épaulée dans les moments difficiles.
Je suis arrivée au bout ! Merci également pour ta bonne humeur. On est trop fortes !
Merci à tous les [Link] qui ont partagé ma vie au labo. Je remercie le groupe
des [Link] se retrouvant à la pause du midi. Merci en particulier à Raoul, Titouan
et Xabier pour nos sorties culturelles qui nous ont fait reprendre pied dans le monde réel,
Antoine et Mégane pour vos bons conseils, Lisa, Lucien, Lucile, Milan, Mériadec, Emeric,
Grégoire, Maxime. Certains d’entre vous sont devenus de très bons amis et j’espère que
cela continuera après mon départ du labo.
J’aimerais également remercier tous les [Link] et [Link] dont j’ai croisé
la route, Armand, Baptiste, Barbara, Clément, Laetitia, Loïc, Maxime, Michel, Zoé, ...
Une route parfois virtuelle au sein de la famille des [Link] du Discord. Merci
Mathilde et Eva pour cette belle idée ! Toutes ces petites lanternes qui ont éclairé mon
chemin et m’ont permis d’avancer. C’est un endroit où on se sent toujours écouté.es,
[Link] et encouragé.es, j’y ai fait de belles rencontres. Merci aux [Link], à
Erwan, Félix, Floriane, Galaad, Giom, Marie, Marina, Mendeleïev, Nico, Paul et tous les

iii
autres, pour les reprises de chansons, les repas en Salon de Thé, les discussions, les sorties
IRL et surtout merci de rendre cet espace bienveillant.

J’aimerais également remercier mes [Link], beaucoup sont éloigné.es géographique-

ment mais c’est toujours un plaisir de se retrouver. Adrien, Bérenger, Camille, Corinne
(ma partenaire de poterie !), Cyril, Florette, Lélio (quel temps précieux j’ai gagné grâce
à Git et Zotéro !), Lucie, Marie, Maxence, William et Yan.

Puisque, comme dit ma grand-mère, pour bien travailler, il faut bien manger, j’aimerai
remercier mon maraîcher préféré, Bruno, ainsi que sa femme, pour tous les bons légumes
qu’ils font pousser. Mention spéciale pour toutes les variétés de courges que j’ai pu dé-
couvrir ! Merci pour toutes les bonnes idées de recettes et pour le moment de causette du
mardi matin.

Je ne peux pas clore ce chapitre sans remercier tous mes proches de croire en moi et
de me soutenir, ma mère, ma grand-mère, mon père (ici et là haut), Jean-Paul (si on était
partis avec moins d’avance, j’en serai peut être jamais là, qui sait !), Claude et mon neveu
Roman.
Je pense tout spécialement à ma mère et à ma grand-mère. Un grand merci pour votre
amour, et votre soutien qui m’a permis d’avancer dans la voie que j’ai choisie. Vous m’avez
toujours encouragée. Merci maman d’avoir résolu ces carrés magiques et autres casse-têtes
avec moi, c’est le début de ma passion pour les maths ! Merci à ma marraine, Nathalie,
pour ses bons conseils “Une brique après l’autre”.

Enfin je souhaite remercier tout spécialement Joris, mon roc. Merci d’être là et d’avancer
dans la vie à mes côtés.

iv
Soient T ∈ {jours, mois, semaines} et p beaucoup trop proche de 1.

0.4

Au tableau dans 0.8

Au tableau dans 0.2
le bureau de 0.1
mon bureau 0.4
mon directeur
Mr. LATEX a
encore frappé ! 0.2 Discussion
0.3
0.6 0.3 0.6 avec d’autres
0.1 mathématiciens

p l u s t a r d ...
Rédaction
0.4 dans un cahier p
Rédaction
en LATEX
s T

0.6 Terminé ! et
ue

sans erreur lq
(vraiment ..?) que

1−p

La chaîne de Markov de ma thèse en mathématiques...

 TABLE DES MATIÈRES

Plan du manuscrit 1

Introduction 3
1 Processus cinétiques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Système cinétique et physique . . . . . . . . . . . . . . . . . . . . . 3
1.2 Force extérieure aléatoire . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Étude de systèmes stochastiques cinétiques . . . . . . . . . . . . . . 9
2 Dynamique stochastique . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1 Précédents travaux . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Quelques cas particuliers . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Heuristique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Résultats et discussions . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Quelques généralisations du processus directeur . . . . . . . . . . . 31
3 Dynamique stochastique dans un potentiel confinant . . . . . . . . . . . . . 34
3.1 Heuristique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2 Résultats et discussions . . . . . . . . . . . . . . . . . . . . . . . . . 35
4 Perspectives de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1 Généralisation à une force de frottement non linéaire . . . . . . . . 41
4.2 Force aléatoire de Feller . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Généralisation de la force de frottement et du potentiel . . . . . . . 42
4.4 Force de frottement aléatoire . . . . . . . . . . . . . . . . . . . . . . 42

I Stochastic dynamics 43

1 Behavior of a time-inhomogeneous Kolmogorov type diffusion 45

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.2 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.3 Existence and non-explosion of solution . . . . . . . . . . . . . . . . . . . . 50
1.4 Moment estimates of the velocity process . . . . . . . . . . . . . . . . . . . 57

vii
TABLE DES MATIÈRES

1.5 Proof of the asymptotic behavior of the solution . . . . . . . . . . . . . . . 59

1.5.1 Asymptotic behavior in the super-critical regime . . . . . . . . . . . 59
1.5.2 Asymptotic behavior in the critical regime . . . . . . . . . . . . . . 62
1.5.3 Asymptotic behavior in the sub-critical regime . . . . . . . . . . . . 67
1.6 An extension to multiplicative noise . . . . . . . . . . . . . . . . . . . . . . 71
1.6.1 Clock change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
1.6.2 Diffusion coefficient with scaling property . . . . . . . . . . . . . . . 72
1.A Some technical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

2 Behavior of a time-inhomogeneous kinetic Lévy-driven model 79

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.2 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.3 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.4 Moment estimates and non-explosion of the velocity process . . . . . . . . 86
2.5 Proof of the asymptotic behavior of the solution . . . . . . . . . . . . . . . 94
2.5.1 Asymptotic behavior in the super-critical regime . . . . . . . . . . . 95
2.5.2 Asymptotic behavior in the critical regime . . . . . . . . . . . . . . 96
2.5.3 Asymptotic behavior in the sub-critical regime . . . . . . . . . . . . 102
2.6 Extended results in the Lévy case . . . . . . . . . . . . . . . . . . . . . . . 104
2.6.1 Large time behavior of the Lévy driving process . . . . . . . . . . . 105
2.6.2 Proofs of Theorems 2.6.1 and 2.6.2 . . . . . . . . . . . . . . . . . . 110
2.6.3 Lévy driving process with a Brownian component . . . . . . . . . . 116
2.A Some technical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

II Stochastic dynamics in a confining potential 125

Introduction 127

3 Behavior of the non-autonomous system driven by a Brownian motion133

3.1 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3.2 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.3 Moment estimates and non-explosion . . . . . . . . . . . . . . . . . . . . . 136
3.4 Asymptotic behavior of the solution . . . . . . . . . . . . . . . . . . . . . . 137
3.4.1 Convergence of the f.d.d. in the super-critical regime . . . . . . . . 138
3.4.2 Convergence of the f.d.d. in the critical and sub-critical regimes . . 140

viii
 TABLE DES MATIÈRES

3.5 Proof of Corollary 3.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

4 Behavior of the non-autonomous system driven by a Lévy process 145

4.1 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.1.1 Settings on Lévy processes and functional spaces . . . . . . . . . . . 145
4.1.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.2 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.3 Moment estimates and non-explosion . . . . . . . . . . . . . . . . . . . . . 149
4.4 Asymptotic behavior of the solution . . . . . . . . . . . . . . . . . . . . . . 153
4.4.1 Convergence of the f.d.d. in the super-critical regime . . . . . . . . 155
4.4.2 Convergence of the f.d.d. in the critical and sub-critical regime . . . 158
4.5 Proof of Corollary 4.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A Study of the deterministic underlying ODE 165

B Some technical results 168

Bibliography 173

ix
 P LAN DU MANUSCRIT

Dans cette thèse, nous étudions le comportement asymptotique de solutions de sys-

tèmes stochastiques cinétiques inhomogènes en temps de la forme

 F (Vt )

 dVt = dLt − dt − ∇U(Xt ) dt,
tβ

dXt
 = Vt dt.

Dans ce modèle, β est un réel positif et L est un processus de Lévy. La fonction F est une
force de rappel, vérifiant des propriétés d’invariance d’échelle et U représente un poten-
tiel confinant. Ainsi, le processus V représente la vitesse d’une particule de position X,
évoluant dans un potentiel U, et soumise à une force de frottement F . Les interactions
entre la particule et son environnement sont modélisées par le processus L. Notons qu’en
temps long, la force attractive F est modérée par la présence du facteur t−β .
L’enjeu est alors de comprendre comment interagissent les différentes forces, et de montrer
que le processus (Xt/ε , Vt/ε )t>0 , correctement renormalisé, admet une limite en loi explicite
lorsque ε tend vers zéro.

Le manuscrit est découpé en deux parties, la première s’intéresse à l’étude du système

en l’absence de potentiel confinant, i.e. U = 0, la seconde au cas où U est un potentiel
quadratique. Chacune des deux parties est elle-même subdivisée en différents chapitres,
suivant le type de processus directeur L considéré, qu’on suppose être soit un mouvement
brownien, soit un processus de Lévy α-stable.

Dans l’introduction qui suit, nous décrivons brièvement les motivations du modèle et
les résultats existant dans la littérature, en expliquant les différentes approches utilisées.
Dans chacun des cadres considérés, nous présentons alors la synthèse des contributions de
cette thèse. Dans le cas d’une particule non-confinée, les résultats sont détaillés dans la
Section 2.4 (Théorèmes 2.5.1 et 2.5.2), tandis que ceux correspondant au cas d’une partic-
ule confinée dans un potentiel quadratrique sont détaillés dans la Section 3.2 (Théorèmes
3.1.1 et 3.1.2). Nous précisons systématiquement l’approche et les méthodes employées.

1
Plan du manuscrit

Nous concluons l’introduction par quelques perspectives de recherche.

La suite du manuscrit suit le plan suivant.

Partie 1 : Dynamique stochastique

On se place, ici, dans le cadre d’un système cinétique où le processus vitesse V est défini
indépendamment du processus position X, i.e. U = 0.

Chapitre 1 : Cas d’un système dirigé par un mouvement brownien

Ce chapitre est adapté de
â M. Gradinaru and E. Luirard, Asymptotic Behavior for a Time-Inhomogeneous Kol-
mogorov Type Diffusion, 2021, arXiv: 2004.11576

Chapitre 2 : Cas d’un système dirigé par un processus de Lévy α-stable

â M. Gradinaru and E. Luirard, Asymptotic Behavior for a Time-Inhomogeneous
Stochastic Differential Equation Driven by an α-Stable Lévy Process, 2021, arXiv:
2112.07287

Partie 2 : Dynamique stochastique dans un potentiel quadratique

Dans cette partie, les équations vérifiées par les processus vitesse et position sont couplées,
via le potentiel quadratique U.
â T. Cavallazzi and E. Luirard, Asymptotic Behavior of a Time-Inhomogeneous Sto-
chastic System in the Quadratic Potential, in preparation.

Chapitre 3 : Cas d’un système dirigé par un mouvement brownien

Chapitre 4 : Cas d’un système dirigé par un processus de Lévy α-stable

2
 I NTRODUCTION

1 Processus cinétiques
1.1 Système cinétique et physique
Intéressons-nous à une particule, de masse 1, soumise à des forces extérieures. Sup-
posons que la seule force qui s’exerce sur elle soit une force de frottement f . Alors, le
principe fondamental de la dynamique s’écrit

v 0

= −f (t, vt ),
t
(1)
 x0

= vt .
t

Énonçons quelques propriétés sur la force de frottement f . Si la particule n’a pas de

vitesse, il n’y a pas de frottement, ainsi en tout temps t, f (t, 0) = 0. Par ailleurs, no-
tons sgn la fonction signe, avec la convention sgn(0) = 0. Puisqu’une force de frottement
s’oppose au mouvement de la particule, on peut supposer, sans perte de généralité, qu’il
existe une fonction b : R+ × R → R+ telle que, pour tout v ∈ R et t ∈ R+ , on ait
f (t, v) = sgn(v)b(v, t). Il découle de ces propriétés que 0 est un point stable et attractif.

Donnons, dans un premier temps, quelques exemples de forces de frottement classiques

ne dépendant pas de la variable temporelle.
— En mécanique classique, à petite vitesse, la force de frottement est supposée linéaire,
de la forme v 7→ −ρv avec ρ > 0.
— Lorsque la vitesse est plus importante, on suppose, d’après la dynamique des fluides,
que le frottement est quadratique.
De manière plus générale, on pourra choisir b de la forme v 7→ ρ |v|γ , où γ ∈ R+ .
Par ailleurs, il existe également des forces de frottement qui évoluent avec le temps. Citons
quelques exemples.
— Un skieur de fond subit une force de frottement proportionnelle au temps qui passe,
au fur et à mesure que le fart de ses skis s’use.

3
Introduction

— La pierre de curling est soumise à une force de frottement qui diminue avec les
balayages des joueurs.
— La force de frottement d’une fusée qui décolle diminue avec son éloignement à la
Terre. En effet, la pression atmosphérique diminue avec le changement d’altitude et,
avec elle, la force de traînée.
Ainsi, il est intéressant de rajouter une dépendance en temps à la fonction b. Dans la
suite, cette dépendance sera supposée polynomiale, de la forme t−β , où β ∈ R.

Pour résumer, on supposera dorénavant que la force de frottement est de la forme

(t, v) 7→ −ρ sgn(v) |v|γ t−β , où ρ > 0, γ ∈ R+ et β ∈ R sont des paramètres du système.
Par souci de simplicité, on notera F la fonction v 7→ ρ sgn(v) |v|γ .

1.2 Force extérieure aléatoire

En 1908, dans [Lan08], le physicien français Paul Langevin s’intéresse à décrire le
mouvement d’une particule en suspension dans un liquide. En prenant en compte la
mécanique de Newton, il propose une modélisation qui semble être plus réaliste que le
modèle brownien. Pendant un intervalle de temps dt, la particule subit un grand nombre
de chocs microscopiques, ce qui maintient son agitation. Ces interactions entre la particule
et son environnement sont modélisées par une force de perturbation aléatoire. Les chocs
étant supposés indépendants, identiquement distribués, centrés et de carré intégrable,
le théorème central limite (formalisé par le principe d’invariance de Donsker en 1951)
suggère de modéliser cette force par un bruit blanc. Pour être plus précise, en prenant les
notations différentielles et en notant B un mouvement brownien, le système (1) devient
le système dégénéré suivant 
dVt

= dBt − Vt dt,
(2)
dXt

= Vt dt.

C’est la première apparition d’une Équation Différentielle Stochastique (EDS). Les EDS
plus générales sont apparues au milieu du xxe siècle, par les approches différentes mais
complémentaires de Kiyoshi Itô et Wolfgang Döblin. On pourra se référer à [Yor08], pour
plus de détails à ce sujet.
Comme pour l’étude des Équations Différentielles Ordinaires (EDO), le caractère linéaire
de ce système rend son étude plus simple : on connaît une solution explicite de (2), à

4
 Introduction

[Link] © Philippe BOURJON

Figure 1 – Photo d’un jeune Kuhlia mugil.

savoir,  Z t
Vt

 = e−t v0 + es−t dBs ,
0
Z t  (3)
= x0 + v0 (1 − e−t ) + 1 − es−t dBs = x0 + v0 + Bt − Vt .

Xt

0

Remarquons que le processus X est gaussien mais n’est pas un processus de Markov.
Dans [CCM10], les auteurs introduisent le modèle Persistent Turning Walker, inspiré du
mouvement d’une espèce de poisson vivant à la Réunion, le Kuhlia mugil.
La dynamique du poisson est décrite par le système cinétique suivant, faisant intervenir
la position x ∈ R2 , l’angle vitesse θ ∈ R et la courbure κ ∈ R,
 √
dκt



= 2α dBt − κt dt,

dθt = κt dt, (4)



dx = eiθt dt.

t

Le terme aléatoire représente l’exploration de l’environnement par le poisson, tandis que

la force de rappel vers l’origine traduit le fait que le poisson a tendance à reprendre une
trajectoire rectiligne. Remarquons que le couple (κ, θ) est l’équivalent du couple (V, X)
du système (2).
Le comportement asymptotique du système (4) a précédemment été étudié dans [DM08] à
l’aide d’outils propres aux Équations aux Dérivées Partielles (EDP). En effet, ce système
cinétique peut être traduit en une EDP : l’équation de Fokker-Planck. Plus précisèment,
la densité de probabilité p(t, x, θ, κ), si elle existe, vérifie l’équation suivante (aussi appelée
équation de Kolmogorov)

5
Introduction

∂t p + eiθ .∇x p + κ∂θ p − ∂κ (κp) − α2 ∂κκ

2
p = 0, t ≥ 0, x ∈ R, θ ∈ R, κ ∈ R.

Cette densité représente la probabilité de trouver le poisson à l’instant t dans un pe-

tit voisinage de (x, θ, κ). Citons [Ber21, Section 1.3.4 p. 26-29] pour plus de détails sur
l’équation de Fokker-Planck.
Le système d’EDS est étudié d’un point de vue stochastique dans [CCM10]. Les auteurs
montrent par des outils d’analyse stochastique que le poisson atteint un comportement
“stationnaire” proche de celui d’un mouvement brownien. Puisque ce dernier est récur-
rent en dimension 2, on en déduit que le poisson aura tendance à se déplacer dans tout
l’espace, et à revenir en tout point une infinité de fois.

Comme nous l’avons vu précédemment, les forces de frottement ne sont pas nécessaire-
ment linéaires, ni homogènes en temps. Le système (2) pourrait donc prendre la forme
plus générale suivante
sgn(Vt ) |Vt |γ


dVt = dBt − ρ

dt,
tβ (SKEB )

dXt = Vt dt.


L’équation de Fokker-Planck associée s’écrit alors, pour t > 0, x ∈ R, v ∈ R,

" #
sgn(v) |v| 1 2
∂t p(t, x, v) + v∂x p(t, x, v) − ∂v ρ β
p(t, x, v) − ∂vv p(t, x, v) = 0.
t 2

Une trajectoire de la solution de (SKEB ) a été simulée et est représentée dans la Figure 2.

Figure 2 – Exemple de trajectoire des processus vitesse (à gauche) et position (à droite),

dirigés par un mouvement brownien, pour ρ = 1, γ = 1 et β = 2.

6
 Introduction

Figure 3 – Exemple de trajectoire des processus vitesse (à gauche) et position (à droite),

dirigés par un processus de Lévy (1.3)-stable, pour ρ = 1, γ = 1 et β = 2.

Force aléatoire discontinue : les processus de Lévy

En outre, certains travaux s’intéressent à des forces de perturbation qui ne sont pas
nécessairement supposées continues, mais qui ont des sauts de type Lévy. Citons par
exemple [Dit99], où l’évolution du climat, étudiée à partir de la température de calottes
glaciaires du Groenland, est modélisée par un processus de Langevin, dirigé notamment
par un processus α-stable.
Ainsi, une généralisation possible de (SKEB ) est le système suivant, dirigé par un processus
de Lévy L,
sgn(Vt ) |Vt |γ


dVt = dLt − ρ

dt,
tβ (SKEL )

dXt = Vt dt.


Une trajectoire du processus solution est représentée par la Figure 3.

L’équivalent de l’équation de Fokker-Planck pour une force de perturbation de type Lévy
est l’équation de Fokker-Planck fractionnaire, faisant intervenir des dérivées fractionnaires.
Pour plus d’informations à ce sujet, on pourra se référer à [CG00].

Rappelons brièvement la définition et quelques propriétés des processus de Lévy.

Définition 1.1. Soit (Ω, F, (Ft )t≥0 , P) un espace probabilisé filtré. On dit qu’un processus
(Ft )t≥0 -adapté (Lt )t≥0 est un processus de Lévy si
• L0 = 0 p.s,
• t 7→ Lt est presque sûrement continue à droite avec des limites à gauche (càdlàg),
• pour tout 0 ≤ s ≤ t, les accroissements Lt − Ls sont indépendants de Fs et station-

7
Introduction

naires.
De plus, notons S(α, σ, β, µ) la loi stable, où α ∈ (0, 2] est le paramètre de stabilité. Sa
fonction caractéristique est donnée, pour ξ ∈ R, par

πα
   
α
ψ(ξ) := exp iµξ − σ|ξ| (1 − iβ tan sgn(ξ)) ,
2
  1
avec la convention β tan πα
2
= 0, si α = 1. Lorsque Lt − Ls ∼ S(α, (t − s) α , 0, 0), on dit
que L est un processus de Lévy α-stable symétrique.

Remarquons que le mouvement brownien est un processus de Lévy 2-stable. C’est le

seul processus stable symétrique dont les trajectoires sont continues. Puisque pour tout
1 L
c > 0, on a l’invariance (c α Lt )t≥0 = (Lct )t≥0 , les processus de Lévy stables sont aux
processus càdlàg ce qu’est le mouvement brownien aux processus continus.
Par ailleurs, on connaît une écriture explicite pour un processus α-stable.

Proposition 1.2. Soit L un processus de Lévy α-stable, avec α ∈ (0, 2). Sa mesure de
Lévy est donnée par

a+ 1{z>0} + a− 1{z<0}
ν(dz) = dz, avec a+ , a− ≥ 0 et a+ + a− > 0.
|z|1+α

Elle vérifie Z
(1 ∧ z 2 )ν(dz) < +∞.
R∗

De plus, d’après la décomposition d’Itô-Lévy, il existe une mesure aléatoire de Poisson N

et sa mesure compensée N f telles que, pour tout t ≥ 0,

Z t Z


 zN (ds, dz) si α ∈ (0, 1),
∗
Z0 t ZR


 Z tZ
Lt = z N (ds, dz) +
f zN (ds, dz) si α = 1,

 0 Z{0<|z|<1} 0 {|z|≥1}

Z t
si α ∈ (1, 2).
 f(ds, dz)

 zN
0 R∗

Plus α est petit, plus |z|−1−α décroît pour 0 < |z| < 1 et croît pour |z| > 1. Ainsi,
d’après la forme de sa mesure de Lévy, un processus α-stable se déplace essentiellement
par de grands sauts lorsque α est proche de 0, et par de petits sauts lorsque α est proche
de 2 (voir Figure 4). On pourra se référer à [Bas03] pour plus de détails sur les EDS
dirigées par des processus de Lévy.

8
 Introduction

(a) Processus (1/2)-stable (b) Processus (1)-stable (c) Processus (3/2)-stable

symétrique symétrique symétrique

Figure 4 – Exemples de trajectoire de processus de Lévy stables.

1.3 Étude de systèmes stochastiques cinétiques

Face au système (SKEL ), les questions naturelles qui se posent sont semblables à celles
pour un système d’EDO.
— Existe-t-il une solution ? Est-elle unique ?
— La solution, si elle existe, explose-t-elle en temps fini ?
— Quel est le comportement en temps long de cette solution ? Peut-on trouver un
taux de convergence rε tel que rε Xt/ε “converge”, quand ε → 0, vers une limite non
nulle ?
Dans ce dernier point, la notion de convergence est à définir. Soit E un espace de fonctions
à valeurs réelles, muni d’une topologie E. Soient (Zt )t>0 et (Zt∞ )t>0 deux processus de
(ε)
E. Définissons (Zt ) := (rε Zt/ε ) le processus renormalisé avec un taux de convergence
approprié. On peut s’intéresser à diverses notions de convergence. Citons entre autres, de
la plus faible à la plus forte.
— La convergence unidimensionnelle : si Zt converge en loi vers Z ∞ dans R, lorsque
t → +∞. On la notera
Zt =⇒ Z ∞.
t→+∞

— La convergence des marginales de rang fini : si pour tout sous-ensemble fini S de

(ε)
(0, +∞), le vecteur (Zt )t∈S converge en loi vers (Zt∞ )t∈S dans RS , lorsque ε → 0.
On la notera
(ε) f.d.
(Zt )t>0 =⇒ (Zt∞ )t>0 .
ε→0

(ε)
— La convergence du processus : si (Zt )t>0 converge en loi vers (Zt∞ )t>0 dans (E, E),

9
Introduction

lorsque ε → 0. On la notera

(ε)
(Zt )t>0 =⇒ (Zt∞ )t>0 .
ε→0

Remarque 1.3. Il est également possible de s’intéresser à des convergences presque sûres
de type loi du logarithme itéré.

Soit (V, X) la solution de (SKEL ). Dans la suite, on appelle processus renormalisé le

processus (V (ε) , X (ε) ) := (rε,V Vt/ε , rε,X Xt/ε )t≥εt0 , pour des taux de convergence rε,V et rε,V
bien choisis.
Comme expliqué dans [Off14],
Il semble vain de chercher à donner des critères généraux sur le comportement
en temps long de tels processus, nous pouvons cependant nous intéresser à
ceux dont le terme de diffusion et le potentiel possèdent certaines conditions
d’invariance d’échelle.
C’est pourquoi on se concentre dans la suite au système cinétique de la forme (SKEL ).

Remarque 1.4. En reprenant l’approche par l’équation de Fokker-Planck, la convergence

du processus (rε,X Xt/ε , rε,V Vt/ε )t≥εt0 implique, sous condition d’existence d’une densité, la
limite faible suivante, sur l’espace des mesures de probabilités,

lim r−1 r−1 pt/ε (rε,X

−1 −1
x, rε,V v), t > 0, x ∈ R, v ∈ R.
ε→0 ε,X ε,V

De plus, étant donné que la composante aléatoire n’est présente que sur la coordonnée
vitesse, le modèle est dit dégénéré. L’étude du comportement asymptotique de la solution
permet alors de comprendre comment cette force aléatoire se propage sur la coordonnée
position.

Le processus de Kolmogorov. Intéressons-nous tout d’abord à l’exemple fondamental

du processus de Kolmogorov 
dVt

= dBt ,
(5)
dXt

= Vt dt.

Ce processus de Markov a été introduit par Kolmogorov dans [Kol29]. Son générateur
infinitésimal est l’opérateur différentiel d’ordre deux L := 21 ∆ + x∇.
Le processus centré X est gaussien et la convergence de ses lois de dimension finie est

10
 Introduction

caractérisée par la convergence de son noyau de covariance

t s
 
2
K(s, t) := E [Xt Xs ] = s − .
2 6
  
1 3
On remarque donc que la famille de processus ε Vt/ε , ε Xt/ε
2 2 ,ε > 0 est station-
t≥0
naire, de même loi que (V, X).

Remarque 1.5. Le système (SKEB ) peut être vu comme un processus de Kolmogorov

perturbé par la force de frottement f .

Le processus de Langevin. Regardons maintenant du côté du processus de Langevin,

solution de (2). La solution est, elle aussi, explicite, donnée par (3).
Le processus vitesse V est ergodique. On remarque que la limite limt→+∞ E [Vt2 ] existe,
ce qui traduit le fait que l’énergie cinétique moyenne de la particule devient constante.
D’autre part, le processus position X s’exprime à partir d’une intégrale de Wiener, c’est
Zt
à dire une intégrale de la forme h(s) dBs , où l’intégrande h est une fonction détermin-
0
iste. C’est un processus gaussien et à l’aide d’un TCL pour les martingales locales ([JS03,
√ 
Théorème VIII-3.11 p. 473]), on peut montrer que εXt/ε converge en loi vers un
t≥0
mouvement brownien quand ε → 0. On dit alors que le comportement asymptotique est
diffusif.
Remarquons que la composante position perd en régularité à la limite puisque le mouve-
ment brownien est presque sûrement nulle part dérivable. Notons que cette asymptotique
coïncide avec le modèle brownien d’une particule en suspension dans l’eau.

Ces cas particuliers mettent en évidence des comportements nouveaux à la limite :

les taux de convergence diffèrent, les propriétés du processus limite également. Dans la
suite, nous chercherons à déterminer les bons taux de convergence pour lesquels il y a
convergence des processus.

2 Dynamique stochastique
Nous présentons, dans un premier temps, des travaux existant sur le comportement
asymptotique du système (SKEB ).

11
Introduction

2.1 Précédents travaux

Convergence des marginales unidimensionnelles du processus vitesse

La convergence des marginales unidimensionnelles de la diffusion V , solution du sys-

tème (SKEB ) dirigé par un mouvement brownien standard, a été étudiée dans [GO13].
Trois régimes asymptotiques sont mis en évidence en fonction de la position des paramètres
γ et β par rapport à la droite critique 2β = γ + 1.
L’inhomogénéité temporelle de la dérive apporte des difficultés. La stratégie est de se
ramener à une diffusion “presque” homogène, à l’aide d’une transformation d’échelles,
en tirant avantage de la forme du coefficient de dérive et de l’auto-similarité du mou-
vement brownien. Cela repose sur l’observation que si ϕ : [0, t1 ) → [t0 , +∞) est un
C 2 -difféomorphisme, alors  
Z t
dBϕ(s) 
(Wt )t≥0 :=  q
0 ϕ0 (s)
t≥0

reste un mouvement brownien standard.

Les auteurs introduisent donc le processus transformé suivant,

Vϕ(s)
Vs(ϕ) := q , s ∈ [0, t1 ),
ϕ0 (s)

qui est solution de l’EDS

γ+1
ϕ0 (s) 2 (ϕ) γ ϕ00 (s) Vs(ϕ) (ϕ) Vϕ(0)
dVs(ϕ) = dWs − ρ sgn(V s
(ϕ)
) Vs ds − ds, V0 =q .
ϕ(s)β ϕ0 (s) 2 ϕ0 (0)

La tactique étant de faire apparaître un terme homogène en temps dans la dérive, deux
transformations d’échelles apparaissent naturellement :
• la transformation exponentielle, qui vérifie ϕ0e = ϕe et ϕe (0) = t0 ,
2β

• la transformation puissance, qui vérifie ϕ0γ = ϕγγ+1 et ϕγ (0) = t0 .

La transformation exponentielle a aussi été utilisée dans [Bre68] pour transformer un pro-
cessus α-stable en un processus de Markov stationnaire. Dans [AW09], elle est utile pour
se ramener à un processus satisfaisant les hypothèses du théorème de Motoo.

Lorsque 2β = γ + 1, les deux transformations sont identiques et la diffusion V (ϕ) est

homogène en temps. L’étude du processus se ramène alors à celle de sa fonction d’échelle

12
 Introduction

et de sa mesure vitesse. On pourra se reporter par exemple à [KS98, Chap. 5.5] et [Kal02,
Chap. 23] pour plus de détails sur ces deux outils.
Afin de conclure de part et d’autre de la droite critique, le lemme asymptotique suiv-
ant est fondamental, couplé au théorème de comparaison des EDS [RY05, Théorème 3.7
p. 394].

Lemme 2.1 (Gradinaru, Offret). Soient σ, σ∞ , b et b∞ des fonctions continues et B un

mouvement brownien standard. Soient Z et H deux diffusions, uniques solutions faibles
de
dZs = σ(s, Zs ) dBs + b(s, Zs ) ds et dHs = σ∞ (Hs ) dBs + b∞ (Hs ) ds.

Supposons que (Z, H) soit asymptotiquement homogène et ν-ergodique, au sens où

Hs =⇒ ν,
s→+∞

et, les limites suivantes soient satisfaites, uniformément sur tout compact,

lim σ(s, z) = σ∞ (z) et lim b(s, z) = b∞ (z).

s→+∞ s→+∞

Supposons de plus que Z soit borné en probabilité, c’est-à-dire que pour tout ε > 0, il
existe r > 0 tel que
sup P(|Zs | ≥ r) < ε.
s≥0

Alors,
Zs =⇒ ν.
s→+∞

Soient Λρ,γ et Πρ,γ les mesures de probabilités définies par

x2
Λρ,γ (x) := c−1 e− 2 e−Uρ,γ,0 (x) et Πρ,γ (x) := k −1 e−Uρ,γ,0 (x) ,

où c et k sont les constantes de normalisation et

2ρ
Uρ,γ,β (t, x) := |x|γ+1 t−β , t ≥ t0 , x ∈ R.
γ+1

Observons que Λρ,γ peut être vue comme un “mélange” de la loi gaussienne et de la loi
Πρ,γ . Il en découle alors l’un des résultats principaux de [GO13].

13
Introduction

Théorème 2.1.1 (Gradinaru, Offret). Soit V la solution de (SKEB ).

• Si 2β > γ + 1, alors la force de frottement est asymptotiquement négligeable et V se
comporte comme une loi normale

V
√t =⇒ N (0, 1).
t t→+∞

• Si 2β = γ + 1, alors V n’est plus asymptotiquement gaussienne

V
√t =⇒ Λρ,γ .
t t→+∞

• Si 2β < γ + 1, alors le comportement asymptotique de V dépend fortement de la

force de frottement
Vt
β =⇒ Πρ,γ .
t→+∞
t γ+1
Il est naturel de se demander ce qu’il advient de la convergence en tant que processus de
V et de celle du processus position X. Cela fait l’objet du premier chapitre de ce manuscrit.
Notons que la convergence du processus V ne découle pas de manière immédiate de celle
de ses marginales unidimensionnelles. En revanche, la convergence des marginales de rang
fini de V , associée à la tension du processus permet de conclure à la convergence du
processus vitesse.

Modèle cinétique homogène en temps

Dans [FT21], les auteurs étudient le modèle cinétique suivant, dirigé par un mouvement
brownien et dépendant du paramètre ρ > 0,

ρ Vt
dV = dBt − dt,


t
2 1 + Vt2 (6)

dXt

= Vt dt.

Ce modèle fut déjà étudié avec la théorie des EDP dans [NP15], [CNP19] et [LP19].
Puisque les coefficients de (6) ne dépendent pas du temps, les outils propres aux EDS
homogènes en temps sont disponibles. Citons par exemple la fonction d’échelle et la mesure
vitesse. Le résultat principal de [FT21] est le suivant.

Théorème 2.1.2 (Fournier, Tardif). Soit (Vt , Xt ) la solution de (6). Soient (Wt )t≥0 un
(α)
mouvement brownien, (St )t≥0 un processus α-stable symétrique avec α ∈ (0, 2), tel que

14
 Introduction

h i
(α) (δ)
E exp(iξSt ) = exp(−t |ξ|α ), et (Ut )t≥0 un processus de Bessel symétrique de dimen-
sion δ ∈ (0, 1). Les expressions des constantes σρ > 0 sont données dans [FT21, p. 5].
• Si ρ > 5,
1 f.d.
(ε 2 Xt/ε )t≥0 =⇒ (σρ Wt )t≥0 .
ε→0

• Si ρ = 5,
 1  f.d.
|ε log(ε)| 2 Xt/ε =⇒ (σ5 Wt )t≥0 .
t≥0 ε→0

ρ+1
• Si ρ ∈ (1, 5), on pose α = 3

1 f.d. (α)
(ε α Xt/ε )t≥0 =⇒ (σρ St )t≥0 .
ε→0

• Si ρ = 1,
3 f.d. (2/3)
(|ε log(ε)| 2 Xt/ε )t≥0 =⇒ (σ1 St )t≥0 .
ε→0

• Si ρ ∈ (0, 1),
 Z t 
1 3 (1−ρ)
(ε Vt/ε , ε Xt/ε )t≥0
2 2 =⇒ Ut , Us(1−ρ) ds .
ε→0 0 t≥0

1
On observe une transition entre la normalisation du mouvement brownien ε 2 et celle
3
de la primitive d’un mouvement brownien ε 2 .
La preuve de la convergence pour ρ > 5 est classique et repose sur la méthode martingale.
On détaillera celle-ci à la section suivante.
Lorsque ρ ∈ (0, 1), la stratégie de preuve est d’approcher le processus vitesse V par un
processus de Bessel de dimension 1 − ρ, écrit, suivant la représentation de Feller, comme
la fonctionnelle d’un mouvement brownien changé de temps. La limite du couple découle
alors du continuous mapping theorem.
Pour traiter le cas ρ ∈ (1, 5), l’idée est d’écrire le processus X comme la fonctionnelle
d’un mouvement brownien changé de temps. Un résultat de représentation des processus
α-stables (voir [BY87]) permet de conclure.

Cette méthode a été réutilisée et généralisée pour obtenir des théorèmes asympto-
tiques de fonctionnelles additives de V , dans [Bét21]. Notons que ce n’est plus possible
pour un système dont les coefficients dépendent du temps. En revanche, le comportement
asymptotique de certains cas particuliers de (SKEB ) peut être étudié de manière directe.
Décrivons quelques exemples, avant de présenter l’heuristique ainsi que les résultats dans

15
Introduction

le cas général.

2.2 Quelques cas particuliers

Le processus de Kolmogorov

Lorsqu’il n’y a pas de force de frottement, i.e. F = 0, la solution du système (SKEB )

est le processus de Kolmogorov solution de (5). Puisque la famille de processus
 
1 3

ε 2 Vt/ε , ε 2 Xt/ε ,ε > 0
t≥0

est stationnaire, de même loi que (V, X), alors,

 Z t 
1 3 L
 
ε 2 Vt/ε , ε 2 Xt/ε = Bt , Bs ds .
t≥0 0 t≥0

Le processus de Langevin

Le système (2) est homogène en temps et linéaire et le processus V est ergodique. On

note Ve une variable aléatoire ayant pour loi sa mesure invariante. Par ailleurs, l’expression
explicite de X est donnée par (3). En étudiant la convergence de la variation quadratique
de l’intégrale de Wiener définissant X, on conclut à l’aide d’un TCL pour les martingales
locales ([JS03, Théorème VIII-3.11 p. 473]) et en suivant le schéma de preuve [FT21,
Théorème 3], que
 √   
Vt/ε , εXt/ε =⇒ Ve , Bt .
t≥0 ε→0 t≥0

Le cas homogène en temps

Lorsque la force de frottement est indépendante du temps, i.e. β = 0, on a alors à

notre disposition les outils spécifiques aux diffusions homogènes en temps. L’étude de la
mesure vitesse et de la fonction d’échelle de V permet de conclure que le processus vitesse
est ergodique. Notons Ve une variable aléatoire distribuée selon sa mesure invariante. À
défaut de connaître l’expression explicite de X, comme dans le cas linéaire, on peut utiliser
la méthode martingale. Cette méthode a notamment été utilisée dans [CCM10], [EG15]
et [FT21]. Notons L := 21 ∂vv 2
− F ∂v le générateur infinitésimal de l’EDS vérifiée par le
processus V . On commence par résoudre l’équation de Poisson Lg = −Id. En appliquant

16
 Introduction

la formule d’Itô, on obtient alors, pour tout t ≥ 0 et ε > 0,

Z t/ε
rε,X Xt/ε = rε,X [x0 + g(v0 )] − rε,X g(Vt/ε ) + rε,X g 0 (Vs ) dBs . (7)
0

√
Le choix rε,X = ε découle, une nouvelle fois, d’un théorème central limite pour les
martingales locales ([JS03, Théorème VIII-3.22 p. 476]). En effet, notons M (ε) l’intégrale
stochastique de (7). On déduit du théorème ergodique appliqué à V qu’il existe une
constante κ > 0, dépendant de g et de L, telle que

hM (ε) it −→ κt.
ε→0

Ainsi, on a la convergence suivante

√ Z
!
t/ε
0
ε g (Vs ) dBs =⇒ (Bκt )t≥0 .
0 ε→0
t≥0

En outre, en suivant la stratégie de [EG15, Section 3.2], on peut montrer que

√ P
sup εg(Vt/ε ) −→ 0.
t≥0 ε→0

Par conséquent, d’après [Bil99, Théorème 3.1 p. 27], on obtient la convergence de la

composante position,
√ 
εXt/ε =⇒ (Bκt )t≥0 .
t≥0 ε→0

La convergence du couple s’en déduit, comme dans le cas linéaire,

 √ 
Vt/ε , εXt/ε =⇒ (Ve , Bκt )t≥0 .
t≥0 ε→0

Le cas linéaire et inhomogène

Intéressons-nous maintenant au processus de Langevin inhomogène en temps. Il est

possible de résoudre explicitement le système dont il est solution,

Vt

dVt

= dBt − ρ dt,
tβ
dX

= Vt dt.
t

17
Introduction

Le processus V est gaussien. Son expression est explicite et donnée, pour tout t ≥ t0 > 0,
par
t
 Z 1−β 1−β
−ρ t1−β s
e eρ 1−β dBs si β 6= 1,




t0
Vt = v0 +  Z t (8)
1


 ρ sρ dBs si β = 1.
t t0

Son comportement asymptotique découle soit de la théorie des martingales continues, soit
de celle des processus gaussiens.
Supposons tout d’abord que β ≥ 1. Alors, l’étude de la variation quadratique de V
permet de conclure, grâce au théorème central limite pour les martingales locales ([JS03,
Théorème VIII-3.22 p. 476]) que

1
 
ε 2 Vt/ε =⇒ (Bκt )t>0 ,
t≥εt0 ε→0

où κ est définie par 

1 si β > 1,


κ := 1

 si β = 1.
1 + 2ρ

 Z t 
Par ailleurs, définissons la fonction gε : V 7→ Vt , Vs ds . Celle-ci converge vers
εt0 t≥εt0
 Z t 
g : V 7→ Vt , Vs ds . En adaptant la preuve du continuous mapping theorem, il est
0 t>0
possible de montrer que
gε (V (ε) ) =⇒ g(Bκ• ).
ε→0
 1 3

Puisque ε 2 Vt/ε , ε 2 Xt/ε = gε (V (ε) ), on obtient donc la convergence suivante
t≥εt0

 Z t 
1 3
 
ε Vt/ε , ε Xt/ε
2 2 =⇒ Bκt , Bκs ds .
t≥εt0 ε→0 0 t>0

Traitons maintenant le cas β < 1. Rappelons qu’un processus gaussien est caractérisé
par son espérance et son noyau de covariance. Or, on peut montrer que, pour tout
(s, t) ∈ [εt0 , +∞)2 ,
 β β  1
lim Cov ε 2 Vs/ε , ε 2 Vt/ε = (s ∧ t)β 1{s=t} := KV (s, t).
ε→0 2ρ

18
 Introduction

Notons (Vt )t≥0 le processus gaussien centré de noyau de covariance KV . On en déduit

alors que
 β
f.d.

ε 2 Vt/ε =⇒ (Vt )t>0 .
t≥0 ε→0

Si le processus V (ε) était tendu, alors on pourrait conclure à la convergence du processus

β
(ε1+ 2 Xt/ε )t≥εt0 . Le cas β = 0 nous indique que ce n’est pas possible. Puisque la convergence
de V n’a lieu que pour les marginales de rang fini, l’argument utilisé lorsque β ≥ 1 pour
conclure à la convergence du processus position ne fonctionne plus.
En revanche, en reprenant la stratégie de la méthode martingale, on peut appliquer la
formule d’Itô à tβ Vt dans le but de décomposer Xt comme la somme d’une intégrale de
Wiener et d’un terme qui tend vers 0 dans L1 . Le noyau de covariance du processus
1+2β
gaussien converge vers KX : (s, t) 7→ (s∧t) 1+2β
. Ainsi, en définissant (Xt )t≥0 le processus
gaussien centré de noyau de covariance KX , on obtient que

1 f.d.
 
εβ+ 2 Xt/ε =⇒ (Xt )t>0 .
t≥0 ε→0

Les stratégies de preuve pour traiter ces cas particuliers sont propres aux EDS ho-
mogènes en temps ou aux processus gaussiens. Elles ne pourront donc pas être util-
isées dans le cas général. Cependant, elles mettent en lumière des idées qu’il est possible
d’adapter. On tentera notamment de se ramener à des processus homogènes.

2.3 Heuristique
On se place, à partir de maintenant, dans le cadre général d’un processus directeur
qui est un Lévy α-stable, pour α ∈ (0, 2]. Afin d’expliquer l’heuristique du comporte-
ment asymptotique des solutions de (SKEL ), nous commençons par donner les équations
vérifiées par les processus renormalisés. Pour tout t ≥ εt0 , on a
Z t  
(ε) 1−γ β−1 −1 (ε)
Vt = rε,V Vt/ε = rε,V (v0 − Lt0 ) + rε,V Lt/ε − rε,V ε rεγ F rε,V Vs s−β ds, (9)
εt0

Z t
(ε) rε,X
Xt = rε,X Xt/ε = rε,X x0 + Vs/ε ds. (10)
εt0 ε
| {z }
=rε,V

En reprenant l’idée du cas linéaire inhomogène, on peut déduire de la convergence du

processus V (ε) et d’un continuous mapping theorem adapté, la convergence du processus

19
Introduction

X (ε) . Alors, le processus limite serait cinétique, i.e. de la forme (V, 0t V)t>0 .
R

Intéressons-nous donc d’abord à la convergence du processus V (ε) . Seuls les deux derniers
termes de (9) jouent un rôle dans la convergence du processus. Plusieurs situations peu-
vent alors se présenter. Elles correspondent aux deux principales méthodes permettant de
montrer qu’un processus converge.
— La plus classique consiste à montrer la convergence des marginales de rang fini du
processus puis à vérifier un critère de tension.
— La seconde repose sur la décomposition du processus comme somme d’un processus
plus connu qui converge et d’un terme qui tend vers 0 en probabilité, uniformément
sur tout compact.
Grâce à l’auto-similarité des processus de Lévy α-stables, la seconde semble être la plus
1
simple à mettre en œuvre. En effet, pour tout ε > 0, le processus (ε α Lt/ε )t≥0 est un Lévy
α-stable. Il suffit donc de montrer que le dernier terme de (9) tend vers 0 en probabilité,
uniformément sur tout compact. Néanmoins, ce n’est pas toujours vrai, ce terme étant
parfois compensé par le processus directeur. Dans ce cas, on utilisera la méthode classique.

Remarque 2.2. En fait, de manière plus générale, il existe des processus de Lévy tels
que (rε Lt/ε )t≥0 converge, pour une bonne renormalisation rε . On donne des résultats dans
la Proposition 2.8.

On distingue ainsi trois régimes asymptotiques, en fonction de la position des coeffi-

cients de frottement γ et β par rapport à α. On se place sous l’hypothèse α > 1, afin de
β
pouvoir définir le rapport q := γ+α−1 , et ainsi simplifier la disjonction des cas.
— Si q > α1 , la force de rappel induite par le frottement est asymptotiquement néglige-
able. Le processus vitesse-position se comporte comme la diffusion de Kolmogorov.
— En régime critique, i.e. q = α1 , la force de frottement compense la force aléatoire.
La loi limite est la loi cinétique d’un “mélange” entre les lois limites des deux autres
régimes. Elle dépend des paramètres de la force de frottement.
— Lorsque q < α1 , le processus vitesse tend à diverger vers l’infini via la force aléatoire,
mais est très vite rappelé vers 0 par la dérive. Le processus directeur n’est pas
asymptotiquement négligeable. De plus, les marginales de rang fini du processus
vitesse limite dépendent des paramètres de la force de frottement. Comme nous
l’avons vu dans le paragraphe traitant du cas linéaire inhomogène, l’étude du cas
homogène en temps q = 0 semble indiquer qu’il ne peut y avoir de convergence

20
 Introduction

du processus V (ε) , mais seulement de ses marginales de rang fini. On perd donc le
caractère cinétique du processus limite, et avec celui-ci la régularité de sa composante
position.
Lorsque le système est dirigé par un mouvement brownien, on met en évidence le change-
ment de régimes dans la vidéo [Lui22a], ainsi que dans la Figure 5. La Figure 6 représente
le système dirigé par un processus α-stable, dans les différents régimes. Une vidéo est
également disponible ([Lui22b]).

2.4 Résultats et discussions

Nous généralisons dans la première partie de ce manuscrit les résultats de [GO13] à la
convergence du processus (V, X) dirigé par un mouvement brownien standard (Chapitre 1)
puis par un processus de Lévy α-stable (Chapitre 2). Notons que cela répond à une
question ouverte posée dans [Off14].

Existence, unicité et explosion

L’étude de l’existence et de la non-explosion de solutions dans le cadre brownien est

détaillée dans [GO13]. Lorsque le processus directeur n’est plus supposé continu, et que
la dérive est localement lipschitzienne, on déduit des résultats classiques ([Sit05, Lemme
115 p. 78 et Théorème 137 p. 104] et [IW81, Théorème 9.1 p. 231]) l’existence et l’unicité
d’une solution locale. Lorsque la dérive est une fonction hölderienne, on conclut à l’aide
de [CZZ17, Remarque 1.3]. Enfin, pour montrer que le temps d’explosion du processus
solution est presque sûrement infini, on démontre le lemme suivant.
Lemme 2.3. Soit (Yt )t≥t0 la solution locale d’une EDS dirigée par un processus de Lévy.
On définit, pour r ≥ 0, le temps d’arrêt τr := inf{t ≥ t0 , |Yt | ≥ r} et τ∞ := limr→+∞ τr le
temps d’explosion de Y . Supposons qu’il existe deux fonctions mesurables positives φ et b
telles que
(i) φ est croissante et limr→∞ φ(r) = +∞,
(ii) b est à valeurs finies,
(iii) et pour tout t ≥ t0 ,
sup E [φ(|Yt∧τr |)] ≤ b(t).
r≥0

Alors τ∞ = +∞ p.s.
On est donc ramené à étudier les moments du processus vitesse V .

21
Introduction

(a) Processus V (ε) , avec q = 1 (b) Processus X (ε) , avec q = 1

1 1
(c) Processus V (ε) , avec q = 2 (d) Processus X (ε) , avec q = 2

1 1
(e) Processus V (ε) , avec q = 4 (f) Processus X (ε) , avec q = 4

(g) Processus V (ε) , avec q = 0 (h) Processus X (ε) , avec q = 0

Figure 5 – Exemples de trajectoire des processus vitesse (à gauche) et position (à droite)

renormalisés, dans le cas d’un système linéaire dirigé par la même réalisation d’un mou-
vement brownien. [Lui22a] est une vidéo illustrant l’évolution des processus, en fonction
de q.
22
 Introduction

2 2
(a) Processus V (ε) , avec q = α (b) Processus X (ε) , avec q = α

1 1
(c) Processus V (ε) , avec q = α (d) Processus X (ε) , avec q = α

1 1
(e) Processus V (ε) , avec q = 2α (f) Processus X (ε) , avec q = 2α

(g) Processus V (ε) , avec q = 0 (h) Processus X (ε) , avec q = 0

Figure 6 – Exemples de trajectoire des processus vitesse (à gauche) et position (à droite)

renormalisés, dans le cas d’un système linéaire dirigé par la même réalisation d’un (1.3)-
stable. [Lui22b] est une vidéo illustrant l’évolution des processus, en fonction de q.

23
Introduction

Estimée des moments

L’étude des moments occupe un large pan de ce manuscrit. Lorsque V est une dif-
fusion, on montre que V a des moments de tout ordre, comme son processus directeur.
En appliquant la formule d’Itô pour la fonction x 7→ x2 et en utilisant l’attractivité vers
l’origine de la dérive, on conclut avec l’inégalité de Jensen pour les moments d’ordre
κ ∈ [0, 2]. Lorsque κ ≥ 2, on conclut de proche en proche en utilisant le moment d’ordre
κ − 2.
Le cas de processus à sauts est plus complexe. Le calcul des moments d’un processus
de Lévy α-stable repose soit sur son auto-similarité, soit sur la connaissance de sa fonction
caractéristique. Les estimées des moments utilisées dans [LP06] reposent sur la première
méthode, tandis que la deuxième méthode a été utilisée dans [DS15, Théorème 3.1 c)].
Cependant, on ne connaît pas à priori la fonction caractéristique de la solution d’une EDS
dirigée par un processus de Lévy α-stable. D’autre part, un processus α-stable n’admet
que des moments d’ordre κ ∈ [0, α), or la fonction f : x 7→ |x|κ n’est pas C 2 . On ne peut
donc pas appliquer directement la formule d’Itô-Lévy. L’idée est donc d’approcher f par
une suite de fonctions C 2 .
Lorsque α ∈ (0, 1), on fait apparaître le processus de Lévy α-stable L+ s≤t |∆Ls |.
P
t :=
Tandis que si α ∈ (1, 2), l’idée clé pour obtenir une borne optimale des moments est de
1
découper les sauts du processus directeur au niveau ξ 7→ ξ α . Cette méthode a notamment
été utilisée par [CM20]. En fait, si L est un processus de Lévy α-stable, alors par la
décomposition d’Itô-Lévy, pour tout t ≥ t0 ,
Z tZ Z tZ Z tZ
Lt − Lt0 = 1
f(ds, dz) +
zN 1 zN (ds, dz) − 1 zν(dz) ds.
t0 |z|≤ξ α t0 |z|>ξ α t0 |z|>ξ α

(1) (2)
Observons que Lt := tt0 |z|≤ξ α1 z N
R R f(ds, dz) et L Rt R
t := 1
t0 |z|>ξ α zN (ds, dz) vérifient la
propriété d’auto-similarité au temps ξ

(1) L 1 (1) (2) L 1 (2)

Lξ = ξ α L1 et Lξ = ξ α L1 .

κ
On applique alors la formule d’Itô avec la fonction v 7→ (η + v 2 ) 2 . Puis, on estime chacun
des termes en suivant le schéma de preuve de [DS15, Théorème 3.1]. Celui-ci repose sur
l’inégalité de Kunita ([App09, Théorème 4.4.23 p. 265]), une formule de Taylor, et les
propriétés hölderiennes des fonctions puissances. En optimisant selon η, on déduit la
proposition suivante.

24
 Introduction

Proposition 2.4. Soit V la solution de (SKEL ). Pour tous α ∈ (0, 2], γ ∈ R, β ∈ R et

κ ∈ [0, α), il existe une constante Cγ,κ,β,t0 telle que,

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 t α . (11)

Lorsque α = 2, l’inégalité (11) est satisfaite pour tout κ ≥ 0.

Remarque 2.5. Le résultat reste vrai pour toute solution d’EDS avec une dérive attractive
(i.e. telle que pour tout v ∈ R, on ait vF (v, .) ≥ 0) de la forme

dVt = dLt − F (t, Vt ) dt.

Lorsque la dérive n’est plus supposée attractive vers 0, la preuve repose sur l’inégalité
de Jensen et un lemme de Grönwall généralisé (Lemme 1.A.1).

Comportement asymptotique

Énonçons les principaux résultats pour F : v 7→ ρ sgn(v) |v|γ .

β
Dans le cadre brownien, on pose q := et on munit l’espace des fonctions continues
γ+1
C((0, +∞), R) de la topologie uniforme sur tout compact

+∞
1
 
2
X
du : (f, g) ∈ C((0, +∞), R) →
7 n
min 1, sup |f − g| .
n=1 2 [ 1 ,n] n

Théorème 2.5.1. Soient γ ≥ 0 et (Vt , Xt )t≥t0 la solution de (SKEB ).

- Régime sur-critique. Supposons 2q > 1. Soit (Bt )t≥0 un mouvement brownien. Alors,
 Z t 
1 3
 
ε Vt/ε , ε Xt/ε
2 2 =⇒ Bt , Bs ds .
t≥εt0 ε→0 0 t>0

- Régime critique. Supposons 2q = 1. On note Hf le processus stationnaire solution de

l’EDS dirigée par un mouvement brownien (Wt )t≥0

f = dW − H
f
s
 
dHs s ds − F H f ds,
s
2
√ f
telle que H
f
−∞ suit la loi invariante. On note (Vt )t≥0 := ( tHlog(t) )t≥0 .

25
Introduction

Alors,  Z t 
1 3
 
ε Vt/ε , ε Xt/ε
2 2 =⇒ Vt , Vs ds .
t≥εt0 ε→0 0 t>0

- Régime sous-critique. Supposons 2q < 1. On note H c le processus stationnaire, issu

de sa loi invariante Π, solution de l’EDS dirigée par un mouvement brownien (Wt )t≥0
 
c = dW − F H
dH c ds.
s s s

On note (Vt )t≥0 le processus ayant pour marginales la mesure image de la mesure
produit Π⊗d par l’application T (u1 , · · · , ud ) := (t1 q u1 , · · · , td q ud ). Alors,

f.d.
 
εq Vt/ε =⇒ (Vt )t>0 .
t≥εt0 ε→0

De plus, si γ = 1 et β ∈ (− 21 , 1), on définit le processus gaussien (Xt )t≥0 centré et

de noyau de covariance défini par

(s ∧ t)1+2β
∀s, t ≥ 0, K(s, t) := .
ρ2 (1 + 2β)

Alors,
1 f.d.
 
εβ+ 2 Xt/ε =⇒ (Xt )t>0 .
t≥εt0 ε→0

Définissons

Λ := {λ : R+ → R+ , continue croissante t.q. λ(0) = 0, lim λ(t) = +∞}

t→+∞

et 
1




1 si n
≤ t ≤ n,

kn (t) = n + 1 − t si n < t < n + 1,


0 si n + 1 ≤ t.


On munit l’espace des fonctions càdlàg D((0, +∞), R) de la topologie de Skorokhod ds

définie, pour tout (f, g) ∈ D((0, +∞), R)2 par
  
+∞
X 1   λ(t) − λ(s) 
n
1 ∧ inf a, ∃λ ∈ Λ, sup log ≤ a, sup |kn (t) (f ◦ λ(t) − g(t))| ≤ a .
n=1 2  s6 = t t − s t≥ 1 
n

β
Pour α ∈ [1, 2), on rappelle que q = γ+α−1
. Dans ce cas, les régimes dépendent de la

26
 Introduction

position de q par rapport à α1 . On donne les résultats dans le cas général α ∈ (0, 2).

Théorème 2.5.2. Soient α ∈ (0, 2), γ ∈ (1 − α2 , α) et (Vt , Xt )t≥t0 la solution de (SKEL ),

dirigé par un processus de Lévy α-stable.
γ−1
- Régime sur-critique. Supposons β > 1 + α
. Soit (St )t≥0 de même loi que (Lt )t≥0 .
Alors,  Z t 
1 1
1+ α
(ε Vt/ε , ε
α Xt/ε )t≥εt0 =⇒ St , Ss ds .
ε→0 0 t>0

- Régime critique. Supposons β = 1+ γ−1 α

. On note H
f le processus stationnaire solution

de l’EDS dirigée par un processus α-stable R, de même loi que L,

f = dR − H
f
s
 
dHs s ds − F H
f ds,
s
α
1
telle que H−∞ suit la loi invariante. On note (Vt )t≥0 := (t α Hlog(t) )t≥0 .
f f

Alors,  Z t 
1 1
1+ α
(ε Vt/ε , ε
α Xt/ε )t≥εt0 =⇒ Vt , Vs ds .
ε→0 0 t>0

- Régime sous-critique. Supposons α > 1, γ ≥ 1 et β < 1 + γ−1 α

. On note H c le

processus stationnaire, issu de sa loi invariante Π, solution de l’EDS dirigée par un

processus α-stable R, de même loi que L,
 
c = dR − F H
dH c ds.
s s s

On note (Vt )t≥0 le processus ayant pour marginales la mesure image de la mesure
produit Π⊗d par l’application T (u1 , · · · , ud ) := (t1 q u1 , · · · , td q ud ). Alors,

f.d.
 
εq Vt/ε =⇒ (Vt )t>0 .
t≥εt0 ε→0

On représente les résultats sous la forme d’un schéma (Figure 7), qui illustre l’influence
des termes de (9). On observe une transition de phase continue des taux de convergence
rε,V et rε,X : de la normalisation du processus de Langevin (q = 0) à celle du processus
de Kolmogorov (q ≥ α1 ). Le changement de régimes est également visible sur la vidéo de
simulation [Lui22a].

Détaillons les schémas de preuve.

27
Introduction

1 1 1
(ε0 , ε α ) (εq , ?) (ε α , ε1+ α )
• •
0 1 q
α

Limite cinétique
Figure 7 – Transition de phase

On montre, dans un premier temps, à partir d’un continuous mapping theorem adapté,
que si le processus V (ε) converge alors X (ε) converge également. Cela implique que le
processus limite est cinétique.

Régime sur-critique. La preuve est la même, suivant si α = 2 ou α ∈ (0, 2). Par

(ε) 1
auto-similarité, le processus (Lt )t≥0 := (ε α Lt/ε )t≥0 a la même loi que L. De plus, on
montre, à l’aide de l’estimée des moments, en notant r := min(β − 1 + 1−γ α
, α1 ) > 0, que

O (εr ) si β 6= 1 + αγ ,
" # 

(ε) (ε) ε→0
E sup Vt − Lt =  1

εt0 ≤t≤T O ε α |ln(ε)| sinon.


ε→0

Régime critique. S’inspirant des travaux de [GO13], on s’intéresse aux transformations

d’échelles de V . D’après [ST94, Proposition 3.4.1 p. 124], le processus suivant est un
processus α-stable de même loi que L
!
Z t
dLϕ(s)
(Rt )t≥0 := 1 .
0 ϕ0 (s) α t≥0

Ainsi, on obtient un résultat généralisant [GO13, Proposition 2.1].

Proposition 2.6. Soit ϕ : [0, t1 ) → [t0 , +∞) un C 2 -difféomorphisme. Si V est solution

Vϕ
de (SKEL ), alors V (ϕ) := 0 1 est solution de
ϕα
γ−1
ϕ0 (s)1+ α ϕ00 (s) Vs(ϕ) (ϕ) Vϕ(0)
dVs(ϕ) = dRs − F (V s
(ϕ)
) ds − ds, avec V0 = 1 . (12)
ϕ(s)β ϕ0 (s) α ϕ0 (0) α

Inversement, si V (ϕ) est solution de (12), alors V est solution de (SKEL ), avec
Z t
1
Lt − Lt0 := (ϕ0 ◦ ϕ−1 ) α (s) dRϕ−1 (s) .
t0

28
 Introduction

En régime critique, le processus V (e) est solution de l’EDS homogène en temps

Vs(e)   γ
dVs(e) = dRs − ds − ρ sgn Vs(e) Vs(e) ds.
α

On montre dans un premier temps la convergence des marginales de rang fini de V (e) .
Donnons les idées pour les marginales de dimension 2.
Pour cela, on remarque que le processus V (e) admet une probabilité invariante Λ. Et
puisque pour tout (s, t) ∈ [εt0 , +∞)2 , la différence

t s t
     
ϕ−1
e − ϕ−1
e = log
ε ε s

ne dépend pas de ε, on obtient par stationnarité de V (e) que, pour tout ε → 0,

L
   
(e) (e) (e) (e)
Vlog(s)+log((t0 ε)−1 ) , Vlog(t)+log((t0 ε)−1 ) = Vlog(s) , Vlog(t) .

Par ailleurs, en s’inspirant de [CCM10], on montre le lemme général suivant.

Lemme 2.7. Soit H un processus fortement ergodique de mesure invariante ν, solution

d’une EDS dirigée par un processus de Lévy. Soit φ : [t0 , +∞) → R une fonction continue
telle que lims→+∞ φ(s) = +∞.
Alors, pour tout d ∈ N∗ , toute fonction continue bornée ψ : Rd → R, et pour tout
(t1 , · · · , td ) ∈ [εt0 , +∞)d , on a
       
E ψ Hφ(ε−1 t1 ) , · · · , Hφ(ε−1 td ) H0 = h0 − E ψ Hφ(ε−1 t1 ) , · · · , Hφ(ε−1 td ) H0 ∼ ν −→ 0.
ε→0

Le processus V (e) satisfait les conditions du lemme, par conséquent, pour tout
(s, t) ∈ [εt0 , +∞)2 ,

1 1
   
(e) (e)
ε α Vs/ε , ε α Vt/ε =⇒ T ∗ L Vlog(s) , Vlog(t) ,
ε→0

1 1
où T ∗ µ désigne la mesure image de µ par l’application linéaire T (u, v) := (s α u, t α v).
Dans un second temps, on vérifie le critère de tension de Kolmogorov ([Kal02, Corollaire
16.9 p. 313]) ou d’Aldous ([Bil99, Théorème 16.10 p. 178]), à l’aide de la propriété de
Markov du processus de Lévy directeur et de l’estimée des moments du processus vitesse.

29
Introduction

Régime sous-critique. En suivant [GO13], la stratégie est cette fois d’étudier la trans-
formation d’échelles puissance. D’après la Proposition 2.6, le processus V (q) est alors so-
lution de l’EDS dirigée par un processus α-stable R de même loi que L,
  γ
dVs(q) = dRs − ρ sgn Vs(q) Vs(q) ds − qϕαq−1 (s)Vs(q) ds.

Contrairement au régime critique, l’EDS n’est pas homogène en temps. Cependant, en

mimant l’idée du lemme asymptotique (Lemme 2.1), on montre que V (q) est proche de la
solution de
dHs = dRs − ρ sgn (Hs ) |Hs |γ ds.

Plus précisément, on prouve que, pour tout (s, t) ∈ [εt0 , +∞)2 ,

   
(q) (q) P
Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) − Vϕ−1 (ε−1 s) , Vϕ−1 (ε−1 t) −→ 0.
ε→0

Pour cela, il est important de ne pas majorer trop brutalement les termes dans l’égalité
suivante
   
(q) (q)
d H − V (q) = − F (Ht ) − F (Vt ) dt + qϕαq−1 (t)Vt dt.
t
 
(q)
En effet, c’est le terme − F (Ht ) − F (Vt ) qui permet de conclure.
Lorsque F est linéaire, on reproduit la méthode de variation des constantes. Néanmoins,
lorsque la dérive n’est pas linéaire, il est utile de remarquer que F −1 est γ1 -hölderienne, ce
qui permet de se ramener au cas linéaire.
Dans un second temps, la preuve de la convergence des marginales de rang fini du processus
H (ϕ,ε) := (Hϕ−1 (ε−1 t) )t≥εt0 est inspirée de [CCM10], en travaillant avec le semi-groupe asso-
cié à H. En effet, contrairement au régime critique, on a ici pour tout (s, t) ∈ [εt0 , +∞)2 ,

t s
   
−1
ϕ − ϕ−1 −→ +∞.
ε ε ε→0

Puisque H admet une mesure invariante, on s’attend à ce que les marginales de H (ϕ,ε)
se décorrèlent lorsque ε → 0. Le potentiel de l’EDS vérifiée par H étant convexe, le
processus vérifie une inégalité de Poincaré (voir [Bob99]). On conclut à la convergence des
marginales de rang fini de V (ε) en utilisant un résultat de décroissance exponentielle (voir
[BGL14, Théorème 4.2.5 p. 183]).

30
 Introduction

2.5 Quelques généralisations du processus directeur

Processus de Lévy général

Il est ensuite naturel de se demander ce que devient le comportement asymptotique

du système (SKEL ), s’il est dirigé par un processus de Lévy général. Par exemple, qu’en
est-il si on tronque les grands sauts ? Notons que la propriété d’auto-similarité n’est alors
plus vérifiée.
Plaçons-nous dans un premier temps sous l’hypothèse que le processus directeur est un
processus de Lévy sans partie brownienne. On suppose que sa mesure de Lévy ν vérifie
l’une des deux hypothèses suivantes :

g(z)
ν(z) = 1{z6=0} , où g est une fonction positive mesurable telle que
|z|1+α (H1ν,α )
c+ := lim g(z) ≥ 0, c− := lim g(z) ≥ 0,
z→+∞ z→−∞

ou Z
pour un certain α0 > 1, |z|α0 ν(dz) < +∞. (H2ν,α0 )
|z|≥1

Par exemple, la mesure de Lévy d’un processus α-stable tronqué satisfait (H2ν,α0 ) pour tout
α0 > 1, tandis que celle d’un processus α-stable tempéré vérifie (H1ν,α ) (avec g(z) := e−λ|z| ,
où λ > 0).
On introduit également l’hypothèse suivante :

g(z)
ν(z) = 1{z6=0} , où g est une fonction positive mesurable telle que
|z|1+α
(Hg )
Z +∞
|g(z) − g(−z)|
dx < +∞.
1 zα

En s’inspirant de [RT11, Proposition 1], on obtient le comportement asymptotique du

processus directeur L.

Proposition 2.8. Soit (Lt )t≥0 un processus de Lévy de triplet caractéristique (0, ν, b), par
rapport à la fonction h : z 7→ −1 ∨ (z ∧ 1). Les paramètres ν ∗ et b∗ sont explicités dans la
Proposition 2.6.3.
 
(i) Si ν satisfait (H1ν,α ) avec α = 1, c+ = c− = c et (Hg ). Alors le processus εLt/ε
t≥0
converge vers le processus de Lévy 1-stable L, de triplet caractéristique (0, ν ∗ , b∗ ).
 
(ii) Si ν satisfait (H1ν,α ) avec α ∈ (1, 2), et (Hg ). Si b∗ 6= 0, alors le processus εLt/ε
t≥0

31
Introduction

converge vers le processus déterministe L de triplet (0, 0, b∗ ).

 
(iii) Si ν satisfait (H2ν,α0 ) et b∗ 6= 0, alors le processus εLt/ε converge vers le pro-
t≥0
cessus déterministe L∗ de triplet (0, 0, b∗ ).
 1

(iv) Si α ∈ (0, 1), et ν satisfait (H1ν,α ), alors le processus ε α Lt/ε converge vers le
t≥0
processus de Lévy α-stable L de mesure de Lévy ν ∗ .
 1

(v) Si α ∈ (1, 2), b = 0 et ν vérifie (H1ν,α ), alors le processus ε α Lt/ε converge vers
t≥0
le processus de Lévy L de mesure de Lévy ν ∗ et de centre b∗ .

On déduit de ce résultat, le comportement asymptotique de (V (ε) , X (ε) ) en régime

sur-critique. Le taux de convergence de V (ε) est donné par celui de L(ε) et le processus
 
limite est Lt , 0t Ls ds
R
.
t≥0

Par ailleurs, on montre que si le bruit directeur est un processus de Lévy général, alors
la partie brownienne n’amène pas de régularité. Plus précisément, notons (V (1) , X (1) ) la
solution de
 (1) (1) γ

(1)
sgn(Vt ) Vt (1)
= dLt − ρ Vt0 = v0 ∈ R,


dVt dt,
tβ (13)
dX (1) (1) (1)

= Vt dt, Xt0 = x0 ∈ R,

t

et (V (2) , X (2) ) la solution de

 (2) (2) γ

(2)
sgn(Vt ) Vt (2)
= dLt + dBt − ρ Vt0 = v0 ∈ R,


dVt dt,
tβ (14)
dX (2) (2) (2)

= Vt dt, Xt0 = x0 ∈ R.

t

Alors le processus (V (2) , X (2) ) a le même comportement que (V (1) , X (1) ), ce dernier étant
donné par le Théorème 2.5.2. La preuve de ce résultat repose sur le fait que
 
(1) (2) (1) (2)
rε,V Vt/ε − rε,V Vt/ε , rε,X Xt/ε − rε,X Xt/ε
t≥εt0

converge vers 0 en probabilité, uniformément sur tout compact. Ceci découle de l’inégalité
de Burkholder-Davis-Gundy ([App09, Théorème 4.4.22 p. 263]) et du lemme de Grönwall.

32
 Introduction

Bruit multiplicatif brownien

On peut également se demander ce qu’il advient du comportement asymptotique de

(SKEB ) lorsque le terme de diffusion est plus général, par exemple,

F (Vt )
dVt = σ(t, Vt ) dBt − dt, t ≥ t0 . (SDEσ )
tβ

Cela faisait l’objet d’une ouverture de [Off14]. Deux cas sont étudiés dans la Section 1.6
du Chapitre 1. Dans un premier temps, si le coefficient de diffusion σ ne dépend que du
temps t, cela revient à faire un changement d’horloge, et la martingale locale
Z t/ε
(ε)
Mt := σ(s) dBs
t0

est un processus gaussien centré dont on peut étudier le comportement en temps long.
Notre schéma de preuve en régime sur-critique s’adapte donc. En revanche, pour les
régimes critique et sous-critique, nous faisions appel à un changement de temps et d’espace.
Celui-ci ne permet plus
q
de conclure. En effet, quelque soit le choix de ϕ, le terme de dif-
 
fusion σ ϕ(s), Vs(ϕ) ϕ0 (s) dWs reste inhomogène en temps.

Cela nous amène donc à étudier un deuxième type de coefficient de diffusion. On

suppose qu’il existe une fonction f telle que
!
v
pour tout (s, v) ∈ [0, +∞) × R, σ(s, v) = f √ .
s

Il est alors possible d’utiliser le processus changé de temps et d’espace V (ϕ) pour conclure
au comportement asymptotique de V en suivant le même schéma de preuve que [GO13,
Théorèmes 4.1, 4.6 et 4.9]. Notons qu’on obtient ainsi le comportement asymptotique
unidimensionnel de V . Le comportement du processus X reste un problème ouvert.

33
Introduction

3 Dynamique stochastique dans un potentiel confi-

nant
Intéressons-nous maintenant à une particule se déplaçant dans un récipient dont la
forme est donnée par une fonction U. Son énergie potentielle est égale à U à une constante
près. Cette nouvelle force s’exerçant sur la particule est conservative, rendant le système
hamiltonien. La conservation de l’énergie assure qu’elle dérive (par rapport à la position)
de l’énergie potentielle. Dans la Partie II, nous traitons le cas particulier d’un récipient
2
de la forme U : x 7→ x2 . Plus précisément, nous nous intéressons à l’influence du potentiel
quadratique sur le comportement asymptotique du système hamiltonien amorti suivant

sgn(Vt ) |Vt |γ



 dVt = dLt − dt − Xt dt,
tβ (SKEc )

dXt = Vt dt.


Remarque 3.1. Le système (SKEc ) peut également modéliser une particule accrochée à
un ressort. Elle est soumise à une force de rappel de la forme −kx. Celle-ci dérive de
l’énergie potentielle élastique 21 kx2 .

Contrairement à (SKEL ), ce système est couplé. Il n’est donc plus possible d’étudier
la convergence du processus vitesse V sans étudier celle de X. L’existence et l’unicité
d’une solution pour ces systèmes dégénérés, dirigés par un mouvement brownien, ont
précédemment été étudiées dans [Fed+17], [WZ15], [Zha16] et [HMC18]. Lorsque le pro-
cessus directeur est un processus de Lévy, on pourra citer [Zha14] où les coefficients sont
supposés homogènes en temps, et [MM21] lorsque la dérive dépend du temps.
Le comportement asymptotique de systèmes hamiltoniens perturbés par une force aléa-
toire est traité dans [AK94]. Le phénomène d’amortissement résultant de la force de
frottement b est étudié dans [Wu01].

3.1 Heuristique
Afin de s’affranchir du couplage des deux équations, l’idée est de voir (SKEc ) comme
un système en dimension 2. On note, pour tout t ≥ t0 et v ∈ R,
         
Xt  0 0 1 0 0 0
Zt :=  , St :=   , A :=  , Γ :=  et F (t, v) :=  |v|γ  .
 
Vt Lt −1 0 0 1 sgn(v) β
t

34
 Introduction

Alors, le système (SKEc ) s’écrit

dZt = Γ dSt + AZt dt − F (t, Vt ) dt. (15)

Remarquons que A est la matrice de rotation d’angle π2 et que etA est la matrice de rotation
d’angle −t. On voit ainsi apparaître l’oscillateur harmonique sous-jacent au modèle

v 0

= −xt
t
 x0

= vt .
t

L’oscillateur empêche le processus Z (ε) de converger, car celui-ci possède une com-
posante angulaire. Il est en revanche possible d’étudier la convergence de ses marginales
unidimensionnelles. Cependant, pour espérer obtenir une convergence en tant que pro-
cessus, il est nécessaire de retirer la partie oscillante. C’est pourquoi on s’intéresse au
processus Yt := e−tA (Vt , Xt )T , à la place de Zt .
Comme en l’absence de potentiel, on met en évidence trois régimes asymptotiques, en
fonction de la position des coefficients de frottement γ et β par rapport à α. Cependant,
le processus limite n’est plus cinétique, en effet le potentiel quadratique sous-jacent au
modèle facilite la propagation du bruit sur la composante position. Le taux de convergence
et la limite de la composante position sont donc différents.
Le changement de régimes est illustrée par la Figure 8 dans le cas d’un système dirigé par
un mouvement brownien, et par la Figure 9 lorsque le processus directeur est un processus
de Lévy α-stable.

3.2 Résultats et discussions

Le processus Y est donné par

dYt = e−tA Γ dSt − e−tA F (t, Vt ) dt.

Par conséquent, son étude se ramène à celle d’une particule libre, mais en dimension 2.
La robustesse du schéma de preuve en dimension supérieure nous est utile.

35
Introduction

(a) Processus V (ε) , avec q = 1 (b) Processus X (ε) , avec q = 1

1 1
(c) Processus V (ε) , avec q = 2 (d) Processus X (ε) , avec q = 2

1 1
(e) Processus V (ε) , avec q = 4 (f) Processus X (ε) , avec q = 4

(g) Processus V (ε) , avec q = 0 (h) Processus X (ε) , avec q = 0

Figure 8 – Exemples de trajectoire des processus vitesse (à gauche) et position (à droite)

renormalisés, dans le cas d’un système linéaire dans un potentiel quadratique, et dirigé par
la même réalisation d’un mouvement brownien. [Lui22c] est un groupe de vidéos illustrant
l’évolution des processus, en fonction de q.
36
 Introduction

2 2
(a) Processus V (ε) , avec q = α (b) Processus X (ε) , avec q = α

1 1
(c) Processus V (ε) , avec q = α (d) Processus X (ε) , avec q = α

1 1
(e) Processus V (ε) , avec q = 2α (f) Processus X (ε) , avec q = 2α

(g) Processus V (ε) , avec q = 0 (h) Processus X (ε) , avec q = 0

Figure 9 – Exemples de trajectoire des processus vitesse (à gauche) et position (à droite)

renormalisés, dans le cas d’un système linéaire dans un potentiel quadratique, et dirigé
par la même réalisation d’un (1.3)-stable. [Lui22c] est un groupe de vidéos illustrant
l’évolution des processus, en fonction de q.
37
Introduction

Existence, unicité, explosion et estimée des moments

Lorsque la dérive est localement lipschitzienne, on conclut par des résultats classiques
à l’existence d’une solution locale. Lorsque la dérive est seulement hölderienne, l’étude
est plus complexe. Pour conclure, on utilise [HMC18, Théorème 1] et [MM21, Théorème
1]. La non-explosion découle, comme dans le cas d’une particule libre, de l’estimée des
moments, via le Lemme 2.3. Le calcul des estimées s’adapte directement de la preuve des
Chapitres 1 et 2 à la dimension supérieure.

Comportement asymptotique

Citons les principaux résultats obtenus.

Théorème 3.1.1. Soient α = 2 et γ ≥ 0. Soit (Vt , Xt )t≥t0 la solution de (SKEc ). On

β
pose q = . Soit B un mouvement brownien sur R2 .
γ+1
√
- Régime sur-critique. Supposons 2q > 1. Alors le processus ( εYt/ε )t≥εt0 converge
 
vers B 2t .
t>0
√
- Régime critique. Supposons 2q = 1 et γ = 1. Alors le processus ( εYt/ε )t≥εt0 converge
Z t
√
 
vers le processus gaussien √12t s dBs .
0 t>0
- Régime sous-critique. Supposons 2q < 1, γ = 1 et β > 21 . Alors les lois de dimension
 β 
finie de ε 2 Yt/ε convergent vers celles du processus gaussien centré de noyau
t≥εt0
de covariance donné, pour (s, t) ∈ (0, +∞)2 , par K(s, t) = 21 sβ 1{s=t} .

Théorème 3.1.2. Soit α ∈ (0, 2). Supposons que L est un processus de Lévy α-stable
symétrique, d’exposant caractéristique ξ 7→ −a kξkα pour a > 0. Soient γ ∈ (0, α) et
β
(Vt , Xt )t≥t0 la solution de (SKEc ). On pose q := . Soit L un processus de Lévy
γ+α−1
α-stable sur R2 , invariant par rotation, d’exposant caractéristique

a Z 2π
2
ξ ∈ R 7→ −Ce kξkα , avec Ce := |cos(x)|α dx.
2π 0
1
- Régime sur-critique. Supposons α ∈ (1, 2) et αq > 1. Alors le processus (ε α Yt/ε )t≥εt0
converge vers (Lt )t>0 .
1
- Régime critique. Supposons αq = 1 et γ = 1. Alors le processus (ε α Yt/ε )t≥εt0 converge
√
 Z t 
1
vers le processus √t s dLs .
0 t>0

38
 Introduction

- Régime sous-critique. Supposons αq < 1, γ = 1 et β > 21 . Alors, pour tout

 β β

(t1 , · · · , td ) ∈ (0, +∞)d , les marginales de rang fini ε α Yt1 /ε , · · · , ε α Ytd /ε con-
vergent en loi vers la mesure produit µt1 ⊗ · · · ⊗ µtd , où µt est la loi de fonction
caractéristique
2 e
 
α β
ξ 7→ exp − C kξk t .
α
Le potentiel U confine la particule. Cela se lit, par exemple dans le régime sur-critique
brownien, sur la variance du processus limite : Xt se comporte asymptotiquement comme
3
une loi N (0, 2t ) en potentiel confinant, mais comme une loi N (0, t3 ) en l’absence de po-
tentiel.
Notons également que les composantes vitesse et position de Z (i.e. un mélange des pro-
cessus vitesse et position initiaux) deviennent asymptotiquement indépendantes lorsque
le processus directeur est brownien, ce qui n’est pas le cas si ce dernier est un processus
de Lévy α-stable.

Détaillons les schémas de preuve.

Régime sur-critique. On retrouve l’idée utilisée dans le régime sur-critique pour une
particule libre. On définit, pour t ≥ εt0 ,
 
(ε) 1
Z t/ε
1
Z t/ε
− sin(s)
Mt := ε α e−sA Γ dSs = ε α  dLs .
t0 t0 cos(s)

On étudie dans un premier temps la convergence des marginales de rang fini du processus
M (ε) . La théorie des processus gaussiens permet de conclure lorsque α = 2. En effet,
rappelons que la convergence d’un processus gaussien est caractérisée par la convergence
de son espérance et de son noyau de covariance. Lorsque α < 2, la convergence n’a
plus de caractérisation aussi simple. Cependant, M (ε) est une intégrale de Wiener-Lévy,
c’est à dire l’intégrale d’une fonction déterministe contre un processus α-stable, ainsi ses
accroissements sont indépendants. On peut donc se ramener à l’étude de la convergence
des accroissements de l’intégrale, via leur fonction caractéristique.
On vérifie ensuite un critère de tension : le critère de Kolmogorov ([Kal02, Corollaire 16.9
p. 313]) ou d’Aldous ([Bil99, Théorème 16.10 p. 178]).
Enfin, en notant r := min(β −1+ 1−γ α
, α1 ) > 0, on montre à l’aide de l’estimée des moments

39
Introduction

que 
O (εr ) si β 6= 1 + αγ ,
" # 

(ε) (ε) ε→0
E sup Yt − Mt =  1

εt0 ≤t≤T  O ε α |ln(ε)| sinon.
ε→0

Régimes critique et sous-critique. Ces deux régimes sont plus complexes qu’en
l’absence de potentiel confinant. En effet, les composantes vitesse et position étant cou-
plées, une transformation d’échelles ne permet plus de conclure. Pour passer outre cette
difficulté, nous traitons le cas linéaire γ = 1. On voit cette fois-ci apparaître l’oscillateur
harmonique amorti
00 x0 (t)
x (t) + β + x(t) = 0, t ≥ t0 . (16)
t
La stratégie de preuve repose alors sur la théorie des équations différentielles ordinaires.
En effet, on peut étudier le comportement asymptotique d’un système fondamental de
solutions de (16). Notons que si β = 1, le résultat est déjà connu, puisque les fonctions
de Bessel de première et de deuxième espèce d’ordre 0 forment une base de l’espace des
solutions.
Notons R la résolvante associée à (16). En notant f l’ordre de grandeur de la résolvante,
et en définissant, pour tout t ≥ 0,

e−tA Rt (ε)
Z t/ε
Φt := et Mt := vε f (t/ε) Rs−1 Γ dWs ,
f (t) t0

on peut décomposer le processus renormalisé sous la forme

(ε) (ε)
Yt = vε f (t/ε)Φt/ε Rt−1
0
Z0 + Φt/ε M
f .
t

La convergence de Y (ε) découle de la convergence de Φ vers la matrice idendité I2 et

de celle de Mf (ε) , en utilisant le développement asymptotique des solutions de (16). On

conclut à la convergence de M f (ε) avec les mêmes outils que dans le régime sur-critique.

40
 Introduction

4 Perspectives de recherche
On présente dans cette section quelques pistes pouvant ouvrir la voie à de futurs
travaux.

4.1 Généralisation à une force de frottement non linéaire

Dans les Chapitres 1 et 2, nous avons obtenu la convergence du processus position dans
le cas d’une force de frottement linéaire. De plus, la preuve du comportement asymptotique
du processus vitesse-position dans le potentiel quadratique, en régimes critique et sous-
critique, repose sur la linéarité de la force de frottement F .
Le cas d’une force de frottement plus générale fait l’objet de travaux en cours.

4.2 Force aléatoire de Feller

Puisque les accroissements d’un processus de Lévy L sont indépendants et station-
naires, le comportement des sauts de L en t ne dépend ni de t ni de Lt− . En ce sens, les
processus de Feller sont une généralisation des processus de Lévy. En effet, on associe à un
processus de Feller, une famille de triplet {(b(x), Q(x), ν(x, dz)), x ∈ R}. Son générateur
infinitésimal est donné, pour f assez régulière, par

1 Z h i
Af (x) = b(x)∇f (x) + Q(x)∆f (x) + f (x + z) − f (x) − ∇f (x)z1|z|<1 ν(x, dz).
2 z6=0

Ainsi, un processus de Feller se comporte localement comme un processus de Lévy. C’est

pourquoi, on appelle parfois ces processus, des processus de type Lévy. On pourra se
référer, par exemple, à [Küh] pour plus de détails sur ces processus.
On peut donc se demander ce que devient le comportement du système (SKEL ) lorsque le
processus directeur est un processus de Feller général. Par exemple, on peut s’intéresser
à un processus de Feller qui se comporte localement comme un processus α-stable. Pour
une fonction hölderienne α : R → (0, 2], ne s’approchant pas de zéro, le générateur d’un
tel processus est donné par
Z h i cα(x)
Af (x) = f (x + z) − f (x) − ∇f (x)z1{|z|<1} dz.
{z6=0} |z|1+α(x)

Cependant, notons que comme pour les processus de Lévy généraux, nous perdons à priori
la propriété d’auto-similarité, qui est essentielle dans nos preuves.

41
Introduction

4.3 Généralisation de la force de frottement et du potentiel

Plusieurs modèles peuvent être considérés. Une première idée consiste à généraliser
(SKEc ) à une particule se déplaçant dans un paysage plus général. Par exemple, qu’en
est-il s’il y a un double puits du type U : x 7→ x4 − x2 ?
On peut également s’intéresser au système (SKEL ) où la force de frottement dépend de
la position, ou ne possède pas un mais deux points fixes stables.

4.4 Force de frottement aléatoire

Un autre axe de recherche consiste à s’intéresser à une force de frottement aléatoire.
Soient (Bt )t≥0 un mouvement brownien standard et β ∈ R. Soit (W (x))x∈R un mouvement
brownien sur R, indépendant de B. Il est défini, à partir de deux mouvements browniens
standards indépendants W + et W − , par

−
W (x) := Wx+ 1{x>0} + W−x 1{x<0} , x ∈ R.

On se place en environnement aléatoire et on considère la solution formelle du système

1 W 0 (Vt )



 dVt = dBt − dt,
2 tβ (17)

dXt
 = Vt dt.

Le processus vitesse a été étudié dans [Off14]. Heuristiquement, à environnement fixé, le

processus position X semble se comporter comme le système (SKEB ) avec γ = − 12 . Le
cas γ < 0 n’a pas été étudié, mais on s’attend à ce que les résultats soient similaires à
ceux obtenus dans le Chapitre 1.
t t L
Puisque pour tout t ≥ 0, on a l’invariance Tt W (x) := e− 4 W (e 2 x) = W (x), le régime
critique (β = 41 ) fait apparaître l’EDS formelle, qui est, en loi, homogène en temps,

1 x2
dHt = dB̃t − ∂x F (t, Ht ) dt, où F (t, x) := + Tt W (x).
2 2

On devrait donc pouvoir adapter la preuve du régime critique du Chapitre 1 dans ce cas.
Le régime sur-critique, pourtant le plus simple dans ce manuscrit, semble plus complexe
puisque l’écriture (17) est formelle.

42
 Part I

Time-inhomogeneous stochastic
dynamics in a free potential

43
CONTENTS

Contents

1 Behavior of a time-inhomogeneous Kolmogorov type diffusion 45

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.2 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.3 Existence and non-explosion of solution . . . . . . . . . . . . . . . . . . . . 50
1.4 Moment estimates of the velocity process . . . . . . . . . . . . . . . . . . . 57
1.5 Proof of the asymptotic behavior of the solution . . . . . . . . . . . . . . . 59
1.5.1 Asymptotic behavior in the super-critical regime . . . . . . . . . . . 59
1.5.2 Asymptotic behavior in the critical regime . . . . . . . . . . . . . . 62
1.5.3 Asymptotic behavior in the sub-critical regime . . . . . . . . . . . . 67
1.6 An extension to multiplicative noise . . . . . . . . . . . . . . . . . . . . . . 71
1.6.1 Clock change . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
1.6.2 Diffusion coefficient with scaling property . . . . . . . . . . . . . . . 72
1.A Some technical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

2 Behavior of a time-inhomogeneous kinetic Lévy-driven model 79

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.2 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.3 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.4 Moment estimates and non-explosion of the velocity process . . . . . . . . 86
2.5 Proof of the asymptotic behavior of the solution . . . . . . . . . . . . . . . 94
2.5.1 Asymptotic behavior in the super-critical regime . . . . . . . . . . . 95
2.5.2 Asymptotic behavior in the critical regime . . . . . . . . . . . . . . 96
2.5.3 Asymptotic behavior in the sub-critical regime . . . . . . . . . . . . 102
2.6 Extended results in the Lévy case . . . . . . . . . . . . . . . . . . . . . . . 104
2.6.1 Large time behavior of the Lévy driving process . . . . . . . . . . . 105
2.6.2 Proofs of Theorems 2.6.1 and 2.6.2 . . . . . . . . . . . . . . . . . . 110
2.6.3 Lévy driving process with a Brownian component . . . . . . . . . . 116
2.A Some technical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

44
 Chapter 1

B EHAVIOR OF A TIME - INHOMOGENEOUS

KOLMOGOROV TYPE DIFFUSION

Abstract: We study a kinetic stochastic model with a non-linear time-inhomogeneous friction

force and a Brownian-type random force. More precisely, a Kolmogorov type diffusion (V, X) is
considered: here, X is the position of the particle, and V is its velocity. The process V is solution
to a stochastic differential equation driven by a one-dimensional Brownian motion, with a drift
of the form t−β F (v). The function F satisfies some homogeneity condition, and β is a real num-
ber. The behavior in large time of the process (V, X) is proved by using stochastic analysis tools.

Keywords: kinetic stochastic equation; time-inhomogeneous diffusion; explosion time; scaling

transformation; asymptotic distribution; ergodicity.

MSC2010 Subject Classification: Primary 60J60; Secondary 60H10; 60J65; 60F17.

This chapter is based on the article [GL21a], submitted.

45
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

1.1 Introduction
In several domains as fluids dynamics, statistical mechanics, or biology, a number of
models are based on the Fokker-Planck and Langevin equations driven by Brownian mo-
tion, could be non-linear or driven by other random processes. For example, in [CCM10]
the persistent turning walker model was introduced, inspired by the modelling of fish
motion. An associated two-component Kolmogorov type diffusion solves a kinetic system
based on an Ornstein-Uhlenbeck Gaussian process, and the authors study the large-time
behavior of this model by using appropriate tools from stochastic analysis. One of the
natural questions is to understand the behavior in large-time of the solution to the corre-
sponding stochastic differential equation (SDE). Although the tools of partial differential
equations allow us to ask this kind of question, and since these models are probabilistic,
tools based on stochastic processes could be more natural to use.
In the last decade, the asymptotic study of solutions to non-linear Langevin’s type
was the subject of an important number of papers, see [CNP19], [EG15] and [FT21]. For
instance, in [FT21] the following system is studied

ρZ t Z t
Vt = v0 + Bt − F (Vs ) ds and Xt = x0 + Vs ds.
2 0 0

In other words, one considers a particle moving such that its velocity is a diffusion with
an invariant measure behaving like (1 + |v|2 )−ρ/2 , as |v| → +∞. The authors prove that
for large-time, after a suitable rescaling, the position process behaves as a Brownian
motion or other stable processes, following the values of ρ. Results have been extended to
additive functional of V in [Bét21]. It should be noticed that these cited papers use the
standard tools associated with time-homogeneous equations, as invariant measure, scale
function, and speed measure. Several of them will not be available when the drag force
depends explicitly on time. In [GO13], a non-linear SDE driven by a Brownian motion but
having time-inhomogeneous drift coefficient was studied, and its large-time behavior was
described. Moreover, sharp rates of convergence are proved for the 1-dimensional marginal
of the solution. In the present paper, we consider the velocity process as satisfying the
same kind of SDE.

Let us describe our framework: we consider a one-dimensional time-inhomogeneous

stochastic kinetic model driven by a Brownian motion. We denote by (Xt )t>0 the process
describing the position of a particle at time t and having the velocity Vt . The velocity

46
 1.1. Introduction

process (Vt )t>0 is supposed to solve a Brownian-driven SDE with a drag force, varying in
time: Z t
dVt = dBt − b(t, Vt ) dt and Xt = X0 + Vs ds.
0

This system can be viewed as a perturbation of the classical two-component Kolmogorov

diffusion Z t
dVt = dBt and Xt = X0 + Vs ds.
0

In the present paper the drift is supposed to grow slowly to infinity, and it will be supposed
to be of the form t−β F (v), with β ∈ R and F satisfying some homogeneity condition. It
describes a one-dimensional particle subject to a friction force and undergoing many small
random shocks. A natural question is to understand the behavior of the process (V, X)
in large time. More precisely we look for the limit in distribution of (rε,V Vt/ε , rε,X Xt/ε )t ,
as ε → 0, for some rates of convergence. Our results are proved on the product of path
spaces and consequently contain those of [GO13].
1 3
If F = 0, it is not difficult to see that the rescaled position process (ε 2 Vt/ε , ε 2 Xt/ε )t
converges in distribution towards the Kolmogorov diffusion (Bt , 0t Bs ds)t . We prove that
R

this kinetic behavior still holds for sufficiently “small at infinity” drag force. The strategy
to tackle this problem is based on estimates of moments of the velocity process. The main
result can then be extended to the case of a drift being equally weighted in some sense
as the random noise. It either offsets the random noise (critical regime) or swings with it
(sub-critical regime).
As suggested at the beginning of the introduction, other random noises can be considered.
In [GL21b], the case of a Lévy random noise is analyzed. The case of a stochastic system
evolving in a quadratic potential is the purpose of another work (see [CL]).

The organization of the chapter is as follows: in the next section, we introduce no-
tations, and we state our main results. Results about existence and non-explosion of
solutions are stated in Section 1.3. Estimates of the moments of the velocity process are
given in Section 1.4 while the proofs of our main results are presented in Section 1.5. We
deal with an extension to multiplicative noise in the last section.

47
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

1.2 Notations and main results

Let (Bt )t≥0 be a standard Brownian motion, β a real number and F a continuous
function which is supposed to satisfy

for some γ ∈ R, ∀v ∈ R, λ > 0, F (λv) = λγ F (v). (Hγ )

In the following, sgn is the sign function with convention sgn(0) = 0. As an example of
function satisfying (Hγ ) one can keep in mind F : v 7→ sgn(v) |v|γ (see also [GO13]).

Remark 1.2.1. If a function π satisfies (Hγ ), then for all x ∈ R, π(x) = π(sgn(x)) |x|γ .

We consider the following one-dimensional stochastic kinetic model, for t ≥ t0 > 0,

dVt = dBt − t−β F (Vt ) dt, Vt0 = v0 > 0, and dXt = Vt dt, Xt0 = x0 ∈ R. (SKEB )

Most of the convergences take place in the space of continuous functions C((0, +∞), R)
endowed by the uniform topology

+∞
1
 
2
X
du : (f, g) ∈ C((0, +∞), R) →
7 n
min 1, sup |f − g| .
n=1 2 [ 1 ,n] n

(ε)
For a family ((Zt )t>0 )ε>0 of continuous processes, we write

(ε)
(Zt )t>0 =⇒ (Zt )t>0 ,
ε→0

(ε)
if (Zt )t>0 converges in distribution to (Zt )t>0 in C((0, +∞), R), as ε → 0.
We write
(ε) f.d.d.
(Zt )t>0 =⇒ (Zt )t>0 ,
ε→0

(ε)
if for all finite subsets S ⊂ (0, +∞), the vector (Zt )t∈S converges in distribution to
(Zt )t∈S in RS , as ε → 0.
β
Let us state our main results. Set q := .
γ+1
Theorem 1.2.2. Consider γ ≥ 0, and q > 21 . Assume that (Hγ ) is satisfied. Let (Vt , Xt )t≥t0
be the solution to (SKEB ) and (Bt )t≥0 be a standard Brownian motion. Furthermore, if
γ ≥ 1, we suppose that for all v ∈ R, vF (v) ≥ 0.

48
 1.2. Notations and main results

Then,  Z t 
1 3
 
ε Vt/ε , ε Xt/ε
2 2 =⇒ Bt , Bs ds .
t≥εt0 ε→0 0 t≥0

Theorem 1.2.3. Consider γ ≥ 0 and q = 12 . Assume that (Hγ ) is satisfied. Let (Vt , Xt )t≥t0
be the solution to (SKEB ). If γ ≥ 1, we suppose furthermore that for all v ∈ R, vF (v) ≥ 0.
Call Hf the eternal ergodic process, solution to the homogeneous SDE

Hs  
dHs = dWs − ds − F Hs ds,
2

such that the distribution of H −∞ is the invariant measure, where (Wt )t≥0 is again a stan-
f

dard Brownian motion. Setting ΛF,t1 ,··· ,td for the finite dimensional distributions (f.d.d.)
t t≥0 the process whose f.d.d. are T ∗ ΛF,log(t1 ),··· ,log(td ) , the pushforward
f we call (V )
of H,
√ √
measure of ΛF,log(t1 ),··· ,log(td ) by the linear map T (u1 , · · · , ud ) := ( t1 u1 , · · · , td ud ). In-
√ f
deed, we have (Vt )t≥0 = ( tH log(t) )t≥0 .

Then,
 1 3
  Z  t
ε 2 Vt/ε , ε 2 Xt/ε =⇒ Vt , Vs ds .
t≥εt0 ε→0 0 t≥0

Remark 1.2.4. The one-dimensional distribution of (Vt )t≥0 has already been explicitly
computed (see Theorem 4.1 in [GO13]).

Theorem 1.2.5. Consider γ ≥ 1 and q < 21 . Assume that F : v 7→ ρ sgn(v) |v|γ with
ρ > 0. Let (Vt , Xt )t≥t0 be the solution to (SKEB ). Call H
c the ergodic process, solution to

the homogeneous SDE

dHs = dWs − F (Hs ) ds,

where (Wt )t≥0 is a standard Brownian motion. Call ΠF its invariant measure. We call
 
(Vt )t≥0 the process whose f.d.d. are T ∗ Π⊗d F , the pushforward measure of Π⊗d
F by the
q q
linear map T (u1 , · · · , ud ) := (t1 u1 , · · · , td ud ).
Then,
f.d.d.
 
εq Vt/ε =⇒ (Vt )t≥0 .
t≥εt0 ε→0

Moreover, in the linear case (i.e. γ = 1) and if β > − 21 , we define (Xt )t≥0 the centered
Gaussian process with covariance function K(s, t) := (ρ2 (1 + 2β))−1 (s ∧ t)1+2β .
Then, as ε → 0,
1 f.d.d.
 
εβ+ 2 Xt/ε =⇒ (Xt )t≥0 . (1.1)
t≥εt0 ε→0
√
Remark 1.2.6. If β = 0, one can prove using the martingale method, that ( εXt/ε )t≥0
converges towards a Brownian motion. Assume, by way of contradiction, that the process

49
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

(εq Vt/ε )t≥εt0 would converge (i.e. were tight), then by the continuous mapping theorem, the
process (εXt/ε )t≥0 should converge. This is a contradiction with (1.1). Here is why we deal
only with the finite-dimensional convergence of the velocity process.

Remark 1.2.7. Let us give a comment on a generalization to greater dimension. Fix

d ≥ 1. We denote by a dot the usual scalar product and k·k its associated norm. Consider
the following one-dimensional stochastic kinetic model driven by a d-dimensional standard
Brownian motion, defined, for t ≥ t0 > 0, by

Vt kVt kγ−1
dVt = dBt − β
dt, Vt0 = v0 ∈ Rd , and dXt = Vt dt, Xt0 = x0 ∈ Rd . (SKEd )
t

Note that the drift function F : v 7→ v kvkγ−1 satisfies

∀v ∈ R, v · F (v) ≥ 0.

Replacing |·| by k·k, one could prove a similar result as in dimension one.

1.3 Existence and non-explosion of solution

In this section, we will prove the existence of solution to (SKEB ) up to explosion time
and the non explosion of such solution with additional assumption.
In the following, we suppose that γ > −1 and set Ω = C([t0 , +∞)) the set of continuous
functions, that equal +∞ after their (possibly infinite) explosion time. Following the
idea used in [GO13], we first perform a change of time in (SKEB ) in order to produce
at least one time-homogeneous coefficient in the transformed equation. For every C 2 -
diffeomorphism ϕ : [0, t1 ) → [t0 , +∞), let introduce the scaling transformation Φϕ defined,
for ω ∈ Ω, by
ω(ϕ(s))
Φϕ (ω)(s) := q , with s ∈ [0, t1 ).
ϕ0 (s)
The result containing the change of time transformation can be found in Proposition 2.1
p. 187 in [GO13].
Let V be solution to the equation (SKEB ). Thanks
 to Lévy’s characterization theorem of
Z t
dBϕ(s) 
the Brownian motion, (Wt )t≥0 :=  q is a standard Brownian motion. Then,
0 ϕ0 (s)
t≥0

50
 1.3. Existence and non-explosion of solution

by a change of variable t = ϕ(s), one gets

Z tq Z t
F (Vϕ(s) ) 0
Vϕ(t) − Vϕ(0) = ϕ0 (s) dW s − ϕ (s) ds.
0 0 ϕ(s)β

The integration by parts formula yields

  q
Vϕ(s)  ϕ0 (s) ϕ00 (s) Vϕ(s)
d q = dWs − F (V ϕ(s) ) ds − ds.
2ϕ0 (s) ϕ0 (s)
q
ϕ0 (s) ϕ(s)β

As a consequence, we can state the following result in our context.

Proposition 1.3.1. If V is a solution to the equation (SKEB ), then V (ϕ) := Φϕ (V ) is a

solution to
q
ϕ0 (s) q ϕ00 (s) Vs(ϕ) (ϕ) Vϕ(0)
dVs(ϕ) = dWs − F ( ϕ0 (s)Vs(ϕ) ) ds − ds, V0 =q , (1.2)
ϕ(s)β ϕ0 (s) 2 ϕ0 (0)
Z t
dBϕ(s)
where Wt = q .
0 ϕ0 (s)
If V (ϕ) is a solution to (1.2), then Φ−1
ϕ (V
(ϕ)
) is a solution to the equation (SKEB ), where
Z tq
Bt − Bt0 := (ϕ0 ◦ ϕ−1 )(s) dWϕ−1 (s) .
t0
Furthermore, uniqueness in law, pathwise uniqueness or strong existence hold for the
equation (SKEB ) if and only if they hold for the equation (1.2).

In the following, we will use two particular changes of time, depending on which term
of (1.2) should become time-homogeneous.
• The exponential change of time: Setting ϕe : t 7→ t0 et , the exponential scaling
ωt es
transformation is defined by Φe (ω) : s ∈ R+ 7→ √ 0 s , for ω ∈ Ω. Thanks to
t0 e 2
(e)
Proposition 1.3.1, the process V := Φe (V ) satisfies the equation

Vs(e) 1
−β 1
√ s

dVs(e) = dWs − ds − t02 e( 2 −β)s F t0 e 2 Vs(e) ds,
2

where (Wt )t≥0 is a standard Brownian motion.

β
• The power change of time: for q = γ+1 , consider ϕq ∈ C 2 ([0, t1 )) the solution to the
Cauchy problem
ϕ0q = ϕ2q
q , ϕq (0) = t0 .

51
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

 1/(1−2q)
Clearly, ϕq (t) = t1−2q
0 + (1 − 2q)t , when 2q 6= 1, and ϕq = ϕe , when 2q = 1.
The time t1 satisfies t1 = +∞, when 2q ≤ 1, and t1 = t01−2q (2q − 1)−1 , when 2q > 1.
ω(ϕq (s))
The power scaling transformation is defined by Φq (ω) : s ∈ R+ 7→ . The
ϕq (s)q
process V (q) := V (ϕq ) satisfies the equation
q 
dVs(q) = dWs − ϕ−γq
q (s)F ϕ0q (s)Vs(q) ds − qϕ2q−1
q (s)Vs(q) ds, (1.3)

where (Wt )t≥0 is a standard Brownian motion.

We now study the existence and the explosion of the solution to (SKEB ) under the
homogeneity assumption (Hγ ). As a consequence, the power scaling process V (q) satisfies
the equation

dVs(q) = dWs − F (Vs(q) ) ds − qϕ2q−1

q (s)Vs(q) ds, s ∈ [0, t1 ), (1.4)

which can be written, when 2q > 1, as

Vs(q)
dVs(q) = dWs − F (Vs(q) ) ds − δ ds, s ∈ [0, t1 ), (1.5)
t1 − s
q
where δ := . Proposition 3.2, p. 188 in [GO13] can be stated in the present situation.
2q − 1
Proposition 1.3.2. For γ > −1, there exists a pathwise unique strong solution to
(SKEB ), defined up to the explosion time.

Remark 1.3.3. In the linear case (γ = 1), the drift and the diffusion terms are Lips-
chitz and satisfy locally linear growth. The existence and non-explosion of V follow from
Theorem 2.9, p. 289, in [KS98].

Proof. We sketch the proof in our context. Remark first that, since γ > −1, x 7→ |x|γ is
locally integrable. Leaving out the third term on the right-hand side of (1.4), one gets a
time-homogeneous equation:

dHs = dWs − F (Hs ) ds, s ∈ [0, t1 ). (1.6)

By using Proposition 2.2, p. 28, in [CE05], there exists a unique weak solution H to
this time-homogeneous equation (1.6) defined up to the explosion time. Moreover, the
Girsanov transformation induces a linear bijection between weak solutions defined up to

52
 1.3. Existence and non-explosion of solution

the explosion time to equations (1.4) and (1.6). It follows that there exists a unique weak
solution V (q) to equation (1.4). Therefore, by using the bijection induced by the change
of time (Proposition 1.3.1), there exists a unique weak solution V to equation (SKEB ).
Besides, by using Corollary 3.4 and Proposition 3.2, pp. 389-390, in [RY05], pathwise
uniqueness holds for the equation (SKEB ). The conclusion follows (Theorem 1.7, p. 368,
in [RY05]).

Proposition 1.3.4.
• When γ ≤ 1 or for all v ∈ R, vF (v) ≥ 0, then the explosion time of V is a.s.
infinite.
• If 2q > 1, then P(τ∞ = +∞) > 0.
• If γ > 1 and (F (−1), F (1)) ∈ ((0, +∞)) × [0, +∞)) ∪ (R × (−∞, 0)), then we have
P(τ∞ = +∞) < 1, where τ∞ denotes the explosion time of V .

Remark 1.3.5. If 2q < 1 and F (1) = −F (−1) < 0, it follows from Proposition 3.6 p. 9
in [GO13] that the explosion time of V is finite a.s.

Remark 1.3.6. We sketch the proof by using the same argument as in in [GO13].
Actually, one can conclude of the non explosion of the solution by using the moment
estimates computed in Proposition 1.4.1, as in Chapter 2 (see Lemma 2.4.1).

Proof. This proof is inspired by those of Propositions 3.6 and 3.7 in [GO13]. We split
the proof into several steps.
Step 1. Assume first that γ ≤ 1 or vF (v) ≥ 0. We will use a criterion of non explosion
stated in [SV06].
Call Lt the time-inhomogeneous infinitesimal generator of V , then

1 ∂2 F (x) ∂
Lt := 2
− β , t ≥ t0 . (1.7)
2 ∂x t ∂x

Let ψ be a twice continuous differentiable positive function such that

1
for all |x| ≥ 1, ψ(x) = 1 + x2 , for all |x| ≤ , ψ(x) = 1, and ψ ≥ 1, on R.
2
 
Note that ψ does not depend on time. Hence, ∂t + Lt ψ = Lt ψ.
Fix T ≥ t0 and call cT the supremum of the continuous function Lt ψ on [t0 , T ] × [−1, 1].

53
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

Then, for all |x| ≤ 1 and t ∈ [t0 , T ],

Lt ψ(x) ≤ cT ≤ cT ψ(x).

Moreover, for all |x| > 1 and t ∈ [t0 , T ], for a positive constant C,

F (x) ≤ ψ(x)
1

if for all v ∈ R, vF (v) ≥ 0,
Lt ψ(x) = −2x β
+1≤
t 2 max(|F (1)| , |F (−1)|)x2 + 1 ≤ Cψ(x) if γ ≤ 1.


So, by using Theorem 10.2.1, p. 254, in [SV06], we deduce that τ∞ is infinite a.s.
Step 2. In this step, we suppose that 2q > 1. We follow the ideas of the proof of Propo-
sition 3.7, pp. 191-192, in [GO13]. We first show that P(τ∞ = +∞) > 0.
Let V (q) be the pathwise unique strong solution to equation (1.5). And denote by b, the
δ-Brownian bridge, the pathwise unique strong solution to equation

bs
dbs = dWs − δ ds, b0 = x0 , s ∈ [0, t1 ). (1.8)
t1 − s

Note that the equation (1.8) is obtained from (1.5) by omitting the second term on the
(q)
right-hand side. Setting τ∞ for the explosion time of V (q) , then τ∞
(q)
∈ [0, t1 ] ∪ {+∞}
(q)
a.s. and {τ∞ ≥ t1 } = {τ∞ = +∞}. Note that b becomes continuous on [0, t1 ] by setting
bt1 = 0 a.s.
Fix n ≥ 1, and define for all s ∈ [0, t1 ],
n o
Tn := inf s ∈ [0, t1 ), Vs(q) ≥ n , σn := inf{s ∈ [0, t1 ], |bs | ≥ n},

and Z s
1Z s

E(s) := exp −F (bu ) dWu − F (bu )2 du .
0 2 0
Since γ > 1 ≥ 0, we have

1 Z s∧σn 1 Z s∧σn 2γ
     
E exp F (bu )2 du ≤ E exp n max(F (1)2 , F (−1)2 ) du
2 0 2 0
t1 2γ
 
≤ exp n max(F (1)2 , F (−1)2 ) .
2

We observe that Novikov’s condition applies to (Es∧σn )s≥0 . Therefore, by using the Gir-
sanov transformation between b and V (q) , we can write for every n ≥ 1, s ∈ [0, t1 ] and

54
 1.3. Existence and non-explosion of solution

A ∈ Fs ,
h   i h i
(q)
E 1A V•∧Tn 1Tn >s = E 1A (b•∧σn ) E(s ∧ σn )1σn >s .

Letting n → +∞, by dominated convergence theorem and Fatou’s lemma, we obtain

h   i
E 1A V (q) 1τ (q) >s ≥ E [1A (b) E(s)] .
∞

(q)
Hence, P(τ∞ = +∞) = P(τ∞ ≥ t1 ) ≥ E[E(t1 )] > 0.
Step 3. Assume now that γ > 1 and (F (−1), F (1)) ∈ ((0, +∞)) × [0, +∞)) ∪ (R ×
(−∞, 0)). We will show that P(τ∞ = +∞) < 1 when F (1) > 0 and F (−1) > 0. Our
strategy is to apply the criterion for explosion stated at Theorem 10.2.1, p. 254 in [SV06].
Pick T > t0 and choose a ∈ (1, γ). One can choose k ≥ 1 such that a(a − 1)−1 < k(T − t0 ).
−1/2
Introduce the continuous differentiable negative function f1 : x 7→ , and, for µ > 0,
1 + |x|a
the bounded twice continuous differentiable function
 Z x 
G1,µ : x 7→ exp µ f1 (y) dy .
−∞

For all t ∈ [t0 , T ] and all x ∈ R,

1 µ
   
∂t + Lt G1,µ (x) = Lt G1,µ (x) = µG1,µ (x) F (x)t−β |f1 (x)| + f10 (x) + f12 (x)
2 2
1 µ 2
 
−β 0
≥ µG1,µ (x) F (x)T |f1 (x)| + f1 (x) + f1 (x) .
2 2

Since |f1 (x)| ∼ 1

|x|−a , then we have lim F (x) |f1 (x)| = +∞, and using that
|x|→+∞ 2 |x|→+∞
lim f10 (x) = 0, there exists r ≥ 1 such that, for all µ > 0,
|x|→+∞

1
   
−β
∂t + Lt G1,µ (x) ≥ µG1,µ (x) F (x)T |f1 (x)| + f10 (x) ≥ kµG1,µ (x) on [t0 , T ] × [−r, r]c .
2

Moreover, since f12 is bounded away from zero, while |f10 | is bounded on [−r, r] and since
F is non-negative, there exists µ0 , such that,

1 µ0
   
∂t + Lt G1,µ0 (x) ≥ µ0 G1,µ0 (x) f10 (x) + f12 (x) ≥ kµ0 G1,µ0 (x) on [t0 , T ] × [−r, r].
2 2
 
Hence, for all t ∈ [t0 , T ] and all x ∈ R, ∂t + Lt G1,µ0 (x) ≥ kµ0 G1,µ0 (x). Besides, since

55
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

|f1 (x)| ≤ 1 ∧ |x|−a , we get

Z x0   Z  
− f1 (x) dx ≤ 1 ∧ |x|−a dx = a(a − 1)−1 < k(T − t0 ).
−∞ R

Thus, we get a lower bound

G1,µ0 (x0 ) > e−kµ0 (T −t0 ) ≥ e−kµ0 (T −t0 ) sup G1,µ0 (x).
x∈R

Therefore, Theorem 10.2.1, p. 254, in [SV06] applies and V explodes in finite time with
positive probability.
When F (−1) < 0 and F (1) < 0, one can proceed in the same way, using the function
 R 
x 7→ exp µ x+∞ f1 (y) dy instead of G1,µ , in order to get that P(τ∞ = +∞) < 1.
Step 4. It remains to show that P(τ∞ = +∞) < 1 when F (1) < 0 and F (−1) > 0.
As in the previous step, we choose a ∈ (1, γ) and for every T > t0 , we choose again
k ≥ 1 such that a(a − 1)−1 < k(T − t0 ). Moreover, it can be noted that there exists a
continuous differentiable odd function f2 , defined on R, vanishing only at x = 0, such that
|f2 (x)| ≤ 1 ∧ |x|−a , and satisfying

1 1
 
f2 (x) := kx, x ∈ − , , lim |x|γ |f2 (x)| = +∞ and lim f20 (x) = 0.
2k 2k |x|→+∞ |x|→+∞

For µ > 0, we introduce the bounded twice continuous differentiable function

 Z x 
G2,µ : x 7→ exp µ f2 (y) dy .
0

Then for all t ∈ [t0 , T ] and all x ∈ R,

" #
  |F (x)f2 (x)| 1 0 µ
∂t + Lt G2,µ (x) = Lt G2,µ (x) = µG2,µ (x) β
+ f2 (x) + f22 (x)
t 2 2
" γ #
|x| |f2 (x)| 1 0 µ 2
≥ µG2,µ (x) ρ + f2 (x) + f2 (x) ,
tβ 2 2
n o
where ρ = min |F (1)| , |F (−1)| > 0. One can conclude, using the same argument as in
the proof of Proposition 3.7, p. 13, in [GO13].

56
 1.4. Moment estimates of the velocity process

1.4 Moment estimates of the velocity process

In this section, we give estimates for the moment of the velocity process. It will be
useful to control some stochastic terms appearing later.

Proposition 1.4.1. Assume that γ ≥ 0 and β ∈ R. The inequality

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 t 2

holds for
γ+1
• κ ∈ [0, 1], when γ < 1 and β ≥ 2
,
• κ ≥ 0, when for all v ∈ R, vF (v) ≥ 0.
γ+1
If κ ∈ [0, 1], γ < 1 and β < 2
, then
1−β
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 tκ 1−γ .
√
Remark 1.4.2. When −1 < γ < 0, it can be proved that for all t ≥ t0 , E [|Vt |] ≤ Cγ,β,t0 t,
without hypothesis of the positivity of the function v 7→ vF (v).

Proof. Step 1. Assume that γ ≥ 1 and that for all v ∈ R, vF (v) ≥ 0.

Define, for all n ≥ 0, the stopping time Tn := inf{t ≥ t0 , |Vt | ≥ n}. By Itô’s formula, for
all t ≥ t0 , we have
Z t∧Tn Z t∧Tn
2
Vt∧Tn
= v02 + 2Vs dBs − 2s−β Vs F (Vs ) ds + (t ∧ Tn − t0 )
t0 t0
Z t Z t∧Tn
= v02 + 1{s≤Tn } 2Vs dBs − 2s−β Vs F (Vs ) ds + (t ∧ Tn − t0 )
t0 t0
Z t
≤ v02 + 1{s≤Tn } 2Vs dBs + (t − t0 ).
t0

Rt
Since t0 41{s≤Tn } Vs2 ds ≤ 4n2 (t − t0 ) < +∞, taking expectation yields
h i
2
E Vt∧Tn
≤ v02 + (t − t0 ) ≤ Ct0 t.

Set κ ∈ [0, 2], we obtain by Jensen’s inequality that

h iκ  h i κ2 κ
E [|Vt |κ ] ≤ E |Vt |2 2 2
≤ lim inf E Vt∧Tn
≤ Cκ,t0 t 2 . (1.9)
n→+∞

57
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

When κ > 2, the function v 7→ |v|κ is C 2 , so by Itô’s formula, we can write for all t ≥ t0 ,

Z t∧Tn Z t∧Tn
|Vt∧Tn |κ = |v0 |κ + κ sgn(Vs ) |Vs |κ−1 dBs − κs−β |Vs |κ−1 sgn(Vs )F (Vs ) ds
t0 t0
Z t∧Tn
κ(κ − 1)
+ |Vs |κ−2 ds.
t0 2

In addition, using the hypothesis on the sign of F , we have

Z t Z t∧Tn
κ(κ − 1)
|Vt∧Tn |κ ≤ |v0 |κ + 1{s≤Tn } κ sgn(Vs ) |Vs |κ−1 dBs + |Vs |κ−2 ds. (1.10)
t0 t0 2
Rt
We observe that t0 κ2 Vs2κ−2 1{s≤Tn } ds ≤ κ2 n2κ−2 (t − t0 ) < +∞. Taking expectation in
(1.10), we obtain
Z t
κ(κ − 1) h κ−2 i
E [|Vt |κ ] ≤ lim inf E [|Vt∧Tn |κ ] ≤ |v0 |κ + E |Vs | ds.
n→+∞ t0 2
h i
When 0 ≤ κ − 2 ≤ 2, we can upper bound E |Vs |κ−2 by injecting (1.9) and get

κ κ
Z t
κ(κ − 1) κ−2 κ
E [|Vt | ] ≤ |v0 | + Cκ,t0 s 2 ds ≤ Cκ,t0 s 2 .
t0 2

The same method is then applied inductively to prove the inequality for all κ > 2.
Step 2. Assume now that γ ∈ [0, 1[.
Fix κ ∈ [0, 1]. Then Jensen’s inequality yields, for all t ≥ t0 , E [|Vt |κ ] ≤ E [|Vt |]κ , hence it
suffices to verify the inequality only for κ = 1.
Define, for all n ≥ 0, the stopping time Tn := inf{t ≥ t0 , |Vt | ≥ n} and let us recall that
under the hypothesis (Hγ ), there exists a positive constant K, such that |F (v)| ≤ K |v|γ .
We can write, for t ≥ t0 and n ≥ 0,
Z t∧Tn
|Vt∧Tn | ≤ |v0 − Bt0 | + |Bt∧Tn | + s−β |F (Vs∧Tn )| ds
t0
Z t∧Tn
≤ |v0 − Bt0 | + |Bt∧Tn | + Ks−β |Vs∧Tn |γ ds.
t0

58
 1.5. Proof of the asymptotic behavior of the solution

By noting that γ ∈ [0, 1[ and that (Bt2 − t)t≥0 is a martingale, taking expectation we get
Z t
E [|Vt∧Tn |] ≤ E [|v0 − Bt0 |] + E [|Bt∧Tn |] + Ks−β E [|Vs∧Tn |γ ] ds
t0
r h i Z t
≤ E [|v0 − Bt0 |] + 2
E Bt∧Tn
+ Ks−β E [|Vs∧Tn |]γ ds
t0
q Z t
≤ E [|v0 − Bt0 |] + E [t ∧ Tn ] + Ks−β E [|Vs∧Tn |]γ ds
t0
√ Z t
≤ Ct0 t + Ks−β E [|Vs∧Tn |]γ ds.
t0

The function gn : t 7→ E [|Vt∧Tn |] is bounded by n. Applying a Grönwall-type lemma

(Lemma 1.A.1) and Fatou’s lemma, for β 6= 1 and for all t ≥ t0 , we end up with
 1 
√
!
1−γ 1−γ
E [|Vt |] ≤ lim inf E [|Vt∧Tn |] ≤ Cγ Ct0 t + K(t1−β − t01−β ) 
n→+∞ 1−β
√
 t

if β ≥ γ+1
2
,
≤ Cγ,β,t0 1−β
t 1−γ else.


The case β = 1 can be treated similarly.

1.5 Proof of the asymptotic behavior of the solution

This section is devoted to the proof of our main results.

1.5.1 Asymptotic behavior in the super-critical regime

In this section, we assume that γ ≥ 0 and q > 12 .

Proof of Theorem 1.2.2. We split the proof into three steps.

Step 1. We note that it is enough to prove that the process

(ε) √
(Vt )t≥0 := ( εVt/ε )t≥0

converges in distribution to a Brownian motion in the space of continuous functions

C([0, +∞)) endowed by the uniform topology. To see V (ε) as a process of C([0, +∞)),

59
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

(ε) √
let us state for all s ∈ [0, εt0 ], Vs(ε) := Vεt0 = εv0 .
For every ε ∈ (0, 1] and t ≥ εt0 , we can write
Z t
3 3
ε Xt/ε = ε x0 +
2 2 Vs(ε) ds.
εt0

Clearly, the theorem will be proved once we show that gε (V•(ε) ) := (V•(ε) , εt• 0 Vs(ε) ds)
R

converges weakly in C([0, +∞)) endowed by the uniform topology. Here the mapping
 R 
gε : v 7→ vt , εtt 0 vs ds is defined and valued on C((0, +∞)). This mapping is converg-
t≥0  
Rt
ing, as ε → 0, to the continuous mapping g : v 7→ vt , 0 vs ds .
t≥0
We have, for every ε ∈ (0, 1] and t ≥ εt0 ,

(ε) √ √ √ √ Z t/ε
Vt = εVt/ε = ε(v0 − Bt0 ) + εBt/ε − ε F (Vs )s−β ds
t0
√ (ε) β− 12
Z t
= ε(v0 − Bt0 ) + Bt − ε F (Vu/ε )u−β du.
εt0

√
By self-similarity, B (ε) := ( εBt/ε )t≥0 has the same distribution as a standard Brownian
motion.
Assume that the convergence of the rescaled velocity process is proved in the strong way,
that is
(ε) (ε) P
∀T > 0, sup Vt − Bt −→ 0. (1.11)
εt0 ≤t≤T ε→0
 
P
Then it suffices to prove that gε (B (ε) ) =⇒ g(B) and du gε (V (ε) ), gε (B (ε) ) −→ 0, as ε → 0
(see Theorem 3.1 p. 27 in [Bil99]).
On the one hand, the process B (ε) being a Brownian motion and k·k denoting a norm on
R2 , the first convergence follows from

P
∀T > 0, sup kgε (Bt ) − g(Bt )k −→ 0. (1.12)
εt0 ≤t≤T ε→0

Let us prove (1.12). For every εt0 ≤ t ≤ T , we get

Z εt0 Z εt0
kgε (Bt ) − g(Bt )k = Bs ds ≤ |Bs | ds.
0 0

Hence,

√
" # Z εt0 Z εt0
E sup kgε (Bt ) − g(Bt )k ≤ E |Bs | ds ≤ C s ds −→ 0.
εt0 ≤t≤T 0 0 ε→0

60
 1.5. Proof of the asymptotic behavior of the solution

On the other hand, we prove that

(ε) (ε) P
∀T > 0, sup gε (Vt ) − gε (Bt ) −→ 0. (1.13)
εt0 ≤t≤T ε→0

For every εt0 ≤ t ≤ T , using (1.11)

Z t
(ε) (ε) (ε) (ε)
gε (Vt ) − gε (Bt ) = Vt − Bt + Vs(ε) − Bs(ε) ds
εt0
(ε) (ε) P
≤ (1 + T − εt0 ) sup Vt − Bt −→ 0.
εt0 ≤t≤T ε→0

Step 2. Let us prove now (1.11).

Recall that under
!
the hypothesis (Hγ ), there exists a positive constant K, such that
√ γ Vu(ε) γ
( ε) F √ ≤ K Vu(ε) . Modifying the factor in front of the integral part, we get
ε

√ √ Z t √ !
(ε) (γ+1) Vu(ε)
Vt = ε(v0 − Bt0 ) + εBt/ε − ε β− 2
( ε)γ F √ u−β du.
εt0 ε

It follows that, for all T > 0,

√ √
!
(ε) (ε) (γ+1)
Z t
Vu(ε)
sup Vt − Bt ≤ ε |v0 − Bt0 | + ε β− 2 sup ( ε)γ F √ u−β du
εt0 ≤t≤T εt0 ≤t≤T εt0 ε
√ (γ+1)
Z T
γ
≤ ε |v0 − Bt0 | + ε β− 2
K Vu(ε) u−β du.
εt0

Taking the expectation and using moment estimates (Proposition 1.4.1), we obtain, when
β 6= γ2 + 1 and since β > γ+1
2
,
"Z #
(γ+1) T γ (γ+1)
Z T h γi
−β
ε β− 2 E K Vu(ε) u du = εβ− 2 KE Vu(ε) u−β du
εt0 εt0
Z T
(γ+1) γ
≤ εβ− 2 Cγ,β,t0 u 2 −β du
εt0

(γ+1)
β− 2 γ
−β+1
γ
−β+1 √ 
≤C ε T 2 − t0 2
ε −→ 0.
ε→0

61
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

(γ+1)
Hence, setting r = min( 12 , β − 2
) > 0,
" #
(ε) (ε)
E sup Vt − Bt = O (εr ).
εt0 ≤t≤T ε→0

γ
The case β = 2
+ 1 can be treated similarly to get

√
" #
(ε) (ε)
E sup Vt − Bt = O ( ε ln(ε)).
εt0 ≤t≤T ε→0

This concludes the proof.

Remark 1.5.1. One can observe that the only moment in this proof when we need the
condition “γ < 1 or for all v ∈ R, vF (v) ≥ 0” is when we are proving the moment
estimates.

Remark 1.5.2. On can notice that in this proof, we only need an upper bound of the
function F . As a consequence, one can assume the function F to satisfy the following
hypothesis:
|F | ≤ G where G is a positive function satisfying (Hγ ). (H2γ )

This implies that there exists a positive constant K such that, for all v ∈ R, |F (v)| ≤ K |v|γ .
Obviously (H2γ ) is a generalization of (Hγ ). As an example of a function satisfying (H2γ )
(with γ = 0), one can keep in mind F : v 7→ v/(1+v2 ) (see also [FT21]).
The same method cannot be used to conclude to the existence up to the explosion, since the
equation satisfied by the power scaling process (1.3) does not have any time-homogeneous
term. Instead, one gets it by using the exponential change of time process and by consider-
ing G instead of |F |. While, the non-explosion of the solution is ensured by the condition
“γ < 1 or for all v ∈ R, vF (v) ≥ 0”.

1.5.2 Asymptotic behavior in the critical regime

γ+1
Assume in this section that β = 2
.

Proof of Theorem 1.2.3. Step 1. As in the first step of the previous section, it suffices
√
to prove the convergence of the rescaled velocity process ( εVt/ε )t .
Keeping same notations, we prove that gε (V (ε) ) converges in distribution in C([0, +∞))

62
 1.5. Proof of the asymptotic behavior of the solution

to g(V). In order to see V (ε) as a process of C([0, +∞)), let us set for all s ∈ [0, εt0 ],
(ε) √
Vs(ε) := Vεt0 = εv0 . Call Pε , P the distribution of V (ε) and V respectively. Then, using
Portmanteau theorem (see Theorem 2.1 p. 16 in [Bil99]), it suffices to prove that for all
bounded and uniformly continuous function h : C([0, +∞)) × C([0, +∞)) → R,
Z Z
h(gε (ω)) dPε (dω) −→ h(g(ω)) dP (dω).
C([0,+∞))2 ε→0 C([0,+∞))2

Take a bounded and uniformly continuous function h. By assumption, one knows that
Pε =⇒ P , hence, by Problem 4.12 p. 64 in [KS98], it suffices to prove that the uniformly
ε→0
bounded sequence (h ◦ gε ) of continuous functions on C([0, +∞)) converges uniformly on
compact subsets of C([0, +∞)) to the continuous function h ◦ g. Let K be a compact set of
C([0, +∞)). Then, for all ω ∈ K, max[0,εt0 ] |ω| is uniformly bounded by a constant, called
M.
Fix η > 0. By the uniform continuity of h, there exists δ > 0 such that for all ω ∈ K,

du (gε (ω), g(ω)) ≤ δ =⇒ |h ◦ gε (ω), h ◦ g(ω)| ≤ η.

However, there exists ε1 > 0 small enough, such that for all ε ≤ ε1 and for all ω ∈ K,
Z εt0
du (gε (ω), g(ω)) ≤ C ω(s) ds ≤ Cεt0 M ≤ δ.
0

Step 2. We first prove the f.d.d. convergence.

The exponential scaling process V (e) satisfies the time-homogeneous equation

Vs(e)  
dVs(e) = dWs − ds − F Vs(e) ds, (1.14)
2

where (Wt )t≥0 is a standard Brownian motion.

By using Proposition 2.2, p. 28, in [CE05], there exists a unique weak solution H to the
time-homogeneous equation (1.14) defined up to the explosion time. Using the bijection
induced by the exponential change of time (Proposition 1.3.1), we get
!
V t
√ t0 et/2 = (Ht )t≥0 ,
t0 e t≥0

63
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

as solutions to the same SDE, starting at the same point. This can also be written as
!
V
√t = (Hlog(t/t0 ) )t≥t0 .
t t≥t0

So, we have, for all ε > 0, and (t1 , · · · , td ) ∈ [εt0 , +∞)d ,

!
V −1 V −1  
√ ε −1t1 , · · · , √ ε −1td = Hlog(t1 )+log((εt0 )−1 ) , · · · , Hlog(td )+log((εt0 )−1 ) . (1.15)
ε t1 ε td

As in [GO13], the scale function and the speed measure of H are respectively
!
Z x
y2 2
p(x) := exp + sgn(y)F (sgn(y)) |y|γ+1 dy
0 2 γ+1

and !
x2 2
νF (dx) := exp − − sgn(x)F (sgn(x)) |x|γ+1 dx.
2 γ+1
By the ergodic theorem (Theorem 23.15 p. 465 in [Kal02]), H is ΛF -ergodic, where ΛF is
the probability measure associated to νF . Call H f the solution to the time homogeneous

equation (1.14) such that the initial condition H f

−∞ has the distribution ΛF .

For (t1 , · · · , td ) ∈ [εt0 , +∞)d , let ΛF,t1 ,··· ,td := L(H

f ,··· ,H
t1
f ) be the distribution of the
td

vector (Ht1 , · · · , Htd ). Then, for all s ∈ R, ΛF,t1 ,··· ,td = ΛF,t1 +s,··· ,td +s . Indeed, thanks to
f f

the invariance property of ΛF , (H f)

t t∈R and (Ht+s )t∈R satisfy the same SDE, starting at
f

the same distribution. As a consequence, for all ε > 0,

 
L H
f
log(t1 )+log((εt0 )−1 ) , · · · , Hlog(td )+log((εt0 )−1 ) = ΛF,log(t1 ),··· ,log(td ) .
f (1.16)

Moreover, by exponential ergodicity, we can prove that for every continuous and bounded
function ψ : Rd → R,
h  i h  i
E ψ Hlog(t1 /(t0 ε)) , · · · , Hlog(td /(t0 ε)) −E ψ H
f
log(t1 /(t0 ε)) , · · · , Hlog(td /(t0 ε))
f −→ 0.
ε→0
(1.17)
We postpone the proof of this convergence in Step 3.
To conclude this step, gather (1.15), (1.16) and (1.17) to get
!
V −1 V −1
√ ε −1t1 , · · · , √ ε −1td =⇒ ΛF,log(t1 ),··· ,log(td ) .
ε t1 ε td ε→0

64
 1.5. Proof of the asymptotic behavior of the solution

This can be written as

√ √ 
εVt1 /ε , · · · , εVtd /ε =⇒ T ∗ ΛF,log(t1 ),··· ,log(td ) ,
ε→0

where T ∗ ΛF,log(t1 ),··· ,log(td ) is the pushforward of the measure ΛF,log(t1 ),··· ,log(td ) by the linear
√ √
map T (u1 , · · · , ud ) := ( t1 u1 , · · · , td ud ).
Step 3. Let us now prove (1.17).
−1
Pick εt0 ≤ s ≤ t. Set h0 = v0 t0 2 . Actually, we prove a more general result, which will
also be useful in the last regime. The convergence (1.17) will be a direct consequence of
this lemma.

Lemma 1.5.3. Let H be an exponential ergodic process with invariant measure ν, solution
to a SDE driven by a Brownian motion. Pick a continuous function φ : [t0 , +∞) → R
satisfying lims→+∞ φ(s) = +∞.
Then, for all integer d ≥ 1, every continuous and bounded function ψ : Rd → R, all h0 ∈ R
and all (t1 , · · · , td ) ∈ [εt0 , +∞)d ,
       
E ψ Hφ(ε−1 t1 ) , · · · , Hφ(ε−1 td ) H0 = h0 − E ψ Hφ(ε−1 t1 ) , · · · , Hφ(ε−1 td ) H0 ∼ ν −→ 0.
ε→0

Proof. For the sake of clarity, let us give a proof for d = 2. The general case d ≥ 2 is
similar.
Let ψ : R2 → R 
be a continuous and 
bounded function.
We set µε := L Hφ(ε−1 s) H0 = h0 and use the generalized Markov property of solution
to SDEs driven by a Brownian motion (see Corollary 16.9 p. 313 in [Kal02]. This leads to
       
E ψ Hφ(ε−1 s) , Hφ(ε−1 t) H0 = h0 = E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ µε

and, since ΛF is invariant,

       
E ψ Hφ(ε−1 s) , Hφ(ε−1 t) H0 ∼ ν = E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ ν .

Then, we are reduced to prove

       
E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ µε − E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ ν −→ 0.
ε→0

65
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

Hence, setting p(t, x, dy) := Px (Ht ∈ dy) and [Link] V for the total variation norm, we get
       
E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ µε − E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 ∼ ν
Z    
≤ E ψ H0 , Hφ(ε−1 t)−φ(ε−1 s) H0 = y (µε (dy) − ν(dy))
R
Z  
≤ kψk∞ p φ(ε−1 s), h0 , dy − ν(dy)
R
 
≤ kψk∞ kp φ(ε−1 s), h0 , · − νkT V .

We let ε → 0, using the exponential ergodicity of H.

√ 
Step 4. Let us prove now the tightness of the family of distributions V (ε) = εVt/ε
t≥εt0
on every compact interval [m, M ], 0 < m ≤ M . We check the Kolmogorov criterion stated
in Problem 4.11 p. 64 in [KS98].
Take ε0 small enough such that for all ε ≤ ε0 , εt0 ≤ m. Fix m ≤ s ≤ t ≤ M and α > 2.
Recalling that B (ε) is a Brownian motion, using Jensen’s inequality, moment estimates
(Proposition 1.4.1) and the relation β = γ+1
2
, we can write
α#
√ Z
"
h αi h αi t/ε
(ε) (ε) −β
E Vt − Vs(ε) ≤ Cα E Bt − Bsε + Cα E ε F (Vu )u du
s/ε
"Z #
t/ε
α 1− α α−1 α −βα
≤ Cα E [|Bt − Bs | ] + Cα ε 2 (t − s) E |F (Vu )| u du
s/ε
Z t/ε
γα
α 1− α α−1 −βα
≤ Cα E [|Bt−s | ] + Cα ε 2 (t − s) u 2 du
s/ε
α
Z t/ε
1− α α
≤ Cα (t − s) + Cα ε
2 2 (t − s) α−1
u− 2 du
s/ε
α
α−1 1− α α
≤ Cα (t − s) + Cα (t − s)
2 (t 2 − s1− 2 )
α
≤ Cα (t − s) 2 + Cα,m,M (t − s)α−1
α
≤ Cα,m,M (t − s) 2 .

Since α > 2, then α2 > 1 and the upper bound does not depend on ε. Furthermore, by
moment estimates (Proposition 1.4.1),
h i √
sup E Vm(ε) ≤ m < +∞.
ε≤ε0

Conclusion. The previous steps yield weak convergence on every compact set (Theorem

66
 1.5. Proof of the asymptotic behavior of the solution

13.1 p. 139 in [Bil99]). The conclusion follows from Theorem 16.7 p. 174 in [Bil99], since
all processes considered are continuous.

Example 1.5.4. We will see that the limiting process V is more explicit in the linear
case (γ = 1). Choose F (1) = 1, F (−1) = −1, the process H f solution to (1.14) is in
3x2
fact an Ornstein-Uhlenbeck process with invariant measure ΛF (dx) := e− 2 dx. It is a
centered Gaussian process, hence for all s1 , · · · , sd , its f.d.d. ΛF,s1 ,··· ,sd are Gaussian. As
a consequence, knowing the covariance function K is enough to provide the distribution
f is a stationary Ornstein-Uhlenbeck process, K(s, t) = 1 e− 32 |t−s| .
of the process. Since H 3
Hence, the limiting process V having f.d.d T ∗ ΛF,log(t1 ),··· ,log(td ) is a centered Gaussian
2
process with covariance function (s, t) 7→ 13 (s∧t)
s∨t
.

1.5.3 Asymptotic behavior in the sub-critical regime

Assume in this section that β < γ+1 2
and F : v 7→ ρ sgn(v) |v|γ with γ ≥ 1. For
simplicity, we shall write ϕ instead of ϕq .

Proof of Theorem 1.2.5. Step 1. We first prove the f.d.d. convergence of the velocity
(ε)
process (Vt )t≥εt0 := (εq Vt/ε )t≥εt0 . Again we give a proof only for d = 2, since the general
case d ≥ 2 is similar.
The power scaling process V (q) , solution to (1.3) satisfies
 
dVs(q) = dWs − F Vs(q) ds − qϕ2q−1 (s)Vs(q) ds.

We call H the ergodic process solution to the SDE

 
dHs = dWs − F Hs ds, with H0 = h0 := v0 t−q
0 . (1.18)

2ρ γ+1
We denote by ΠF (dx) := e− γ+1 |x| dx its invariant measure. Using the bijection induced
by the power change of time (Proposition 1.3.1), as solutions to the same SDE starting
at the same point, we have, for all ε > 0, and (s, t) ∈ [εt0 , +∞)2 ,

Vε−1 s Vε−1 t
   
(q) (q)
εq q , εq q = Vϕ−1 (ε−1 s) , Vϕ−1 (ε−1 t) .
s t

Using Theorem 3.1 p. 27 in [Bil99], it suffices to prove that for all (s, t) ∈ [εt0 , +∞)2 ,

67
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

   
(q) (q)
• Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) − Vϕ−1 (ε−1 s) , Vϕ−1 (ε−1 t) −→ 0.
ε→0
 
• Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) =⇒ ΠF ⊗ ΠF .
ε→0
 
(q) 2

Step 2. We prove that E Ht − −→ 0. Vt
t→+∞
We have
   
(q) (q)
d H − V (q) = − F (Ht ) − F (Vt ) dt + qϕ2q−1 (t)Vt dt.
t

By straightforward differentiation, we can write

 2     
(q) (q) (q) (q)
d H − V (q) = −2 F (Ht ) − F (Vt ) Ht − Vt dt + 2qϕ2q−1 (t)Vt Ht − Vt dt.
t
(1.19)
We set  
(q) 2

g(t) := E Ht − Vt , t ≥ 0.

Taking expectation in (1.19), we get

h  i h  i
(q) (q) (q) (q)
g 0 (t) = −2E F (Ht ) − F (Vt ) Ht − Vt + 2qϕ2q−1 (t)E Vt Ht − Vt .

Since γ ≥ 1, the function F −1 is γ1 -Hölder, therefore there exists Cγ > 0 such that,
 
(q) 1+γ
 h  i
(q) (q)
g 0 (t) ≤ −Cγ E Ht − Vt + 2qϕ2q−1 (t)E Vt Ht − Vt .

Then, by Jensen’s inequality, since γ ≥ 1,

γ+1 h
(q)

(q)
i
g 0 (t) ≤ −Cγ g(t) 2 + 2qϕ2q−1 (t)E Vt Ht − Vt .

Using Cauchy-Schwarz inequality and moment estimates (Proposition 1.4.1), we have

γ+1 1
q
g 0 (t) ≤ −Cγ g(t) 2 + C |q| ϕq− 2 (t) g(t), g(0) = 0.

1
Note that since 2q < 1, then ϕq− 2 (t) −→ 0, therefore the conclusion follows from
t→+∞ 2 
(q)
Lemma 1.A.3. Besides, for all t ≥ εt0 , E Hϕ−1 (ε−1 t) − Vϕ−1 (ε−1 t) = g (ϕ−1 (ε−1 t)) −→ 0.
ε→0
2
Step 3. Pick (s, t) ∈ [εt0 , +∞) . We prove that the solution H to (1.18) satisfies
 
Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) =⇒ ΠF ⊗ ΠF . (1.20)
ε→0

68
 1.5. Proof of the asymptotic behavior of the solution

Observe that

t1−2q − s1−2q
ϕ−1 (ε−1 t) − ϕ−1 (ε−1 s) = −→ +∞. (1.21)
ε1−2q ε→0

By Lemma 1.5.3, for every continuous and bounded function ψ, we can write
       
E ψ Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) H0 = h0 − E ψ Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) H0 ∼ ΠF −→ 0.
ε→0

Hence, it suffices to prove that for every bounded continuous functions f, g : R → R, the
following convergence holds
     
lim E f Hϕ−1 (ε−1 s) g Hϕ−1 (ε−1 t) H0 ∼ ΠF = ΠF (f )ΠF (g).
ε→0

The following reasoning is inspired from the proof of Lemma 3.2 p. 7-8 in [CCM10]. Since
H0 is starting from the invariant measure, up to considering f − ΠF (f ) and g − ΠF (g), we
can assume that f and g have zero ΠF -mean. We call (Pt )t≥0 the semigroup of H, then
we get, by invariance property of ΠF ,
      Z  
E f Hϕ−1 (ε−1 s) g Hϕ−1 (ε−1 t) H0 ∼ ΠF = Pϕ−1 (ε−1 s) f Pϕ−1 (ε−1 t)−ϕ−1 (ε−1 s) g dΠF
Z
= f Pϕ−1 (ε−1 t)−ϕ−1 (ε−1 s) g dΠF .

1+γ
Note that U : v 7→ |v|1+γ is a convex function, thus a λ-Poincaré inequality holds for the
process H (see [Bob99] p. 1904). This implies the exponential decay of the variance (see
Theorem 4.2.5 p. 183 in [BGL14]), i.e. there exists a constant C > 0 such that, since ΠF
is a probability measure,
Z
f Pϕ−1 (ε−1 t)−ϕ−1 (ε−1 s) g dΠF ≤ f Pϕ−1 (ε−1 t)−ϕ−1 (ε−1 s) g
2

≤ kf k∞ Pϕ−1 (ε−1 t)−ϕ−1 (ε−1 s) g

2
−λ(ϕ−1 (ε−1 t)−ϕ−1 (ε−1 s))
≤ C kf k∞ kgk∞ e .

We deduce (1.20) from (1.21).

Step 4. We prove the convergence of the f.d.d. of the position process.
(ε) 1
We set (Xt )t≥εt0 := (εβ+ 2 Xt/ε )t≥εt0 . Take γ = 1 and β ∈ (− 12 , 1). Pick t ≥ εt0 . By Itô’s

69
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

formula applied to tβ Vt , we get

Z t/ε Z t/ε
1−β
(ε) β+ 21 (ε) β+ 12 β+ 21
ρXt =ε (tβ0 v0 + x0 ) − ε 2 tβ
Vt +ε β
s dBs + ε βsβ−1 Vs ds.
t0 t0

Since β > − 21 , the first term converges to 0 in probability as ε → 0. Moreover, by Itô’s

formula, for all t ≥ t0 ,
d h 2i h i
E Vt = −2ρs−β E Vs2 + 1.
dt
Hence, by comparison theorem for ordinary differential equation,
!
h i t1−β Z t
s1−β
E Vt2 ≤ exp(−2ρ ) v02 + exp(2ρ ) ds .
1−β t0 1−β

Using an integration by parts, we deduce that there exists a positive constant C such
that, for all t ≥ t0 ,
h i
E Vt2 ≤ Ctβ .

As a consequence, we obtain
" Z t/ε # Z t/ε
1−β 1−β
β+ 21 β+ 12
h i
(ε) (ε)
E −ε 2 tβ Vt +ε βs β−1
Vs ds ≤ε 2
β
t E Vt +ε βsβ−1 E [|Vs |] ds
t0 t0
1 3β 1−β 3β 1 3β
≤ Cε 2 t 2 + Cε 2 t 2 − Cεβ+ 2 t02 −→ 0.
ε→0

(ε) 1 R t/ε β
It remains to study the centered Gaussian process Mt := εβ+ 2 t0 s dB s. By Itô’s
isometry and since β > − 12 , for all εt0 ≤ s ≤ t, we can write

(ε)
Z s/ε
s1+2β
Cov(Ms(ε) , Mt ) =ε 2β+1 2β
u ds ∼ .
t0 ε→0 1 + 2β

Since the convergence of centered Gaussian processes is characterized by the convergence

of their covariance function, the conclusion follows from Theorem 3.1 p. 27 in [Bil99].

70
 1.6. An extension to multiplicative noise

1.6 An extension to multiplicative noise

Let σ be a continuous function on [0, +∞) × R and define for ρ > 0, the function
F : v 7→ ρ sgn(v) |v|γ . Consider the one-dimensional SDE defined, for t ≥ t0 , by

dVt = σ(t, Vt ) dBt − F (Vt ) dt. (SDEσ )

We assume in the following that there exists a unique solution to (SDEσ ). Notice that
the (SDEσ ) is an extension to the one studied in [GO13].

1.6.1 Clock change

If the diffusion function σ depends only on the time parameter, i.e. σ(t, v) = σ(t),
then the diffusion term is a Brownian motion with a “clock change”. Furthermore, in
the super-critical regime, the asymptotic behavior depends on the one of the centered
Gaussian process
Z t/ε
(ε)
Mt := σ(s) dBs .
t0

/ L1 (+∞) such that σ 2 ∼+∞ f , then

More precisely, if there exists f ∈
Z (s∧t)/ε
(ε)
K (ε) (s, t) := Cov(Ms(ε) , Mt ) ∼ f (s) ds.
ε→0 t0

Example 1.6.1. Assume that f is a bounded T -periodic continuous function. One can
keep in mind the square of cosinus or sinus functions. Setting bxc for the integral part of
a real x, we have
s ∧ t Z t0 +T
 
εK (ε) (s, t) ∼ f (u) du.
ε→0 T t0
√
Consequently, ( εMtε )t≥εt0 converges in f.d.d. to the centered Gaussian process with co-
j kR
t0 +T
variance function K(s, t) := s∧tT t0 f (u) du.
j k
For example, if f ∈ {cos2 , sin2 }, then K(s, t) = s∧t
2π
π. Moment estimates can be com-
puted as in Section 1.4, since the diffusion coefficient is bounded. The tightness cri-
terion checking follows the same line as in Section 1.5.2. As a conclusion, the process
√ 
εVt/ε converges to the centered Gaussian process with covariance function K.
t≥εt0

In both critical and sub-critical regimes, the changed-of-time process is inefficient to

conclude. Indeed, the general diffusion coefficient does not satisfy a scaling transformation
and thus the changed-of-time process get a time-inhomogeneous diffusion term.

71
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

1.6.2 Diffusion coefficient with scaling property

Let us switch now to another appropriate diffusion coefficient. The scaling transfor-
mation Φϕ , applied to an SDE with multiplicative noise, leads to the following result.

Proposition 1.6.2. Let ϕ : [0, t1 ) → [t0 , ∞) be a C 2 -diffeomorphism. If V is a solution

to the equation (SDEσ ), then V (ϕ) := Φϕ (V ) is a solution to
γ+1
ϕ0 (s) 2 ϕ00 (s) Vs(ϕ) Vϕ(0)
 q 
(ϕ)
dVs(ϕ) =σ ϕ(s), Vs(ϕ) ϕ0 (s) dWs − β
F (Vs
(ϕ)
) ds− 0
ds, V0 = q ,
ϕ(s) ϕ (s) 2 ϕ0 (0)
(1.22)
Z t
dBϕ(s)
where Wt := q .
0 ϕ0 (s)
If V (ϕ) is a solution to the equation (1.22), then Φ−1
ϕ (V
(ϕ)
) is a solution to the equation
Z tq
(SDEσ ), where Bt − Bt0 := (ϕ0 ◦ ϕ−1 )(s) dWϕ−1 (s) .
t0
Furthermore, uniqueness in law, pathwise uniqueness or strong existence hold for the
equation (SDEσ ) if and only if they hold for the equation (1.2).

Pick a continuous function f and suppose that

!
v
for all (s, v) ∈ [0, +∞) × R, σ(s, v) = f √ .
s

We suppose that the following usual condition holds: there exists a positive increasing
function h such that

∀(x, y) ∈ R2 , |f (x) − f (y)| ≤ h(|x − y|),

Z (1.23)
and h−2 (u) du = +∞.
0+

Furthermore, assume that the function f satisfies the conditions

Z 0 Z x Z +∞ ! Z x !
s s
exp 2
ds dx = exp 2
ds dx = +∞,
−∞ 0 f (s) 0 0 f (s)
! (H1f )
1 Z x
s
and x 7→ 2 exp − 2
ds ∈ L1 (R).
f (x) 0 f (s)

72
 1.6. An extension to multiplicative noise

We introduce another assumption on f , which will sometimes be imposed in the sequel

Z x ! !
Z 0
2F (s) Z +∞ Z x
2F (s)
exp 2
ds dx = exp ds dx = +∞,
−∞ 0 f (s) 0 0 f 2 (s)
! (H2f )
1 Z x
2F (s)
and x 7→ 2 exp − ds ∈ L1 (R).
f (x) 0 f 2 (s)

As an example of a function satisfying (H1f ) one can keep in mind f : x 7→ |x|a with
a ∈ [0, 21 ). Moreover, the function f : x 7→ |x|a with a ∈ [0, 1+γ
2
∧ 21 ) satisfies (H1f ) and
(H2f ).

• On the one hand, the transformation associated to the exponential change of time
ϕe : t 7→ t0 et satisfies

γ+1
−β ( γ+1 −β)s Vs(e)
dVs(e) = f (Vs(e) ) dWs − t0 2 e 2 F (Vs(e) ) ds − ds. (1.24)
2

• On the other hand, the transformation associated to the power change of time

1
ϕq : t 7→ (t1−2q
0 + (1 − 2q)t) 1−2q

satisfies
dVs(q) = f (Vs(q) ) dWs − F (Vs(q) ) ds − qϕ2q−1
q Vs(q) ds. (1.25)

Remark 1.6.3. Following the same lines as in the proof of Proposition 1.3.2, one can
prove that there exists a pathwise unique strong solution to (SDEσ ), defined up to the
explosion time. Hence, by Proposition 1.6.2, there exists a pathwise unique strong solution
to (1.24) and (1.25).

We introduce the following probability distributions, where k and k ± denotes normaliza-

tion constants, !
k+ Z x
−2s
µ(dx) := 2 exp dx,
f (x) 0 f 2 (s)
!
k Z x
−2F (s) − 2s
Λ(dx) := 2 exp dx,
f (x) 0 f 2 (s)
and
k−
!
Z x
−2F (s)
Π(dx) := 2 exp dx.
f (x) 0 f 2 (s)

73
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

Theorem 1.6.4. Consider γ ≥ 0. Let (Vt )t≥t0 be the solution to (SDEσ ) such that for
 
all (s, v) ∈ [0, +∞) × R, σ(s, v) = f √vs . Assume that (H1f ) is satisfied.
(i) Assume that 2q > 1. Then,
V
√t =⇒ µ.
t t→+∞

(ii) Assume that 2q = 1. Then,

V
√t =⇒ Λ.
t t→+∞

(iii) Assume that 2q > 1 and that (H2f ) is satisfied. Then,

Vt
=⇒ Π.
tq t→+∞

Proof. In the sequel, the ergodic property of time-homogeneous processes follows from
Theorem 23.15 p. 465, in [Kal02] by studying their scale function and speed measure.
(i) Assume that 2q > 1. Following the proof of Theorem 4.6 in [GO13], we prove that
the asymptotic behavior of V is related to the one of the process solution to

Hs
dHs = f (Hs ) dWs − ds.
2

It is equivalent to prove
(e)
Vt =⇒ µ.

Thanks to (H1f ), the process H is ergodic. Moreover, define for v ∈ R,

v
b− (v) := −ρ |v|γ 1v≥0 − ,
2
v
b+ (v) := ρ |v|γ 1v≤0 − ,
2
and denote by V ± the pathwise unique strong solution to equations

(e)
dVs± = f (Vs± ) dBs + b± (Vs± ) ds, V0± = V0 .

By using (H1f ), V + and V − are ergodic processes, and are thus bounded in prob-
ability. Using a comparison theorem (see Theorem 1.3 in [Yam73]), we prove that,
(e)
for all t ≥ 0, Vt− ≤ Vt ≤ Vt+ a.s. This implies that V (e) is bounded in probability.
The conclusion follows from the asymptotic lemma (Lemma 4.5 in [GO13]).

74
 1.A. Some technical results

(ii) Assume that 2q = 1. Let V (e) be the solution to the time-homogeneous SDE (1.24).
Thanks to (H1f ), the process V (e) is ergodic. The conclusion is a direct consequence
of the ergodic theorem.
(iii) Assume that 2q < 1. As in the proof of Theorem 4.9 in [GO13], it is equivalent to
prove that
(q)
Vt =⇒ Π. (1.26)
t→+∞

Denote by Z the ergodic process solution to

dZs = f (Zs ) dWs − F (Zs ) ds. (1.27)

Thanks to (H2f ), Z is ergodic and satisfies (1.26). Let u > 0 be such that

sup |q| φ2q−1 (t) ≤ 1.

t≥u

Introduce Z ± the pathwise unique strong solution to

dZs± = f (Zs± ) dWs − F (Zs± ) ds ± Zs± 1±Zs± ≥0 ds, Zu± = Vu(q) .

Using a comparison theorem (see Theorem 1.3 in [Yam73]), we obtain that, for all
(q)
t ≥ u, Zt− ≤ Vt ≤ Zt+ a.s. By using (H2f ), Z + and Z − are ergodic processes
and thus V (q) is bounded in probability. The conclusion follows again from the
asymptotic lemma (Lemma 4.5 in [GO13]).

1.A Some technical results

Let us state and prove a Grönwall-type lemma which has been used to get moment
estimates.

Lemma 1.A.1 (Grönwall-type lemma). Fix r ∈ [0, 1) and t0 ∈ R. Assume that g is a non-
negative real-valued function, b is a positive function and a is a differentiable real-valued
function. Moreover, suppose that the function bg r is continuous. If
Z t
∀t ≥ t0 , g(t) ≤ a(t) + b(s)g(s)r ds, (1.28)
t0

75
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

then,  
 Z t  1
1 1−r
∀t ≥ t0 , g(t) ≤ 2 1−r a(t) + (1 − r) b(s) ds .
t0

Proof. For t ≥ t0 , since r ≥ 0,

 Z t r
r r
g(t) ≤ a(t) + b(s)g(s) ds ,
t0

then, multiplying by b(t) > 0,

 Z t r
b(t)g(t)r ≤ b(t) a(t) + b(s)g(s)r ds .
t0

Now, let us make appear the derivative of H

 Z t r
0 r 0 r
a (t) + b(t)g(t) ≤ a (t) + b(t) a(t) + b(s)g(s) ds ,
t0

that is

a0 (t) + b(t)g(t)r a0 (t) a0 (t)

 r ≤ b(t) +  r ≤ b(t) + .
a(t)r
Rt Rt
a(t) + t0 b(s)g(s)r ds a(t) + t0 b(s)g(s)r ds

Integrating, since r 6= 1, we obtain

" 1−r #
1 Z t
r 1−r 1 h i Z t
a(t) + b(s)g(s) ds − a(t0 ) ≤ a(t)1−r − a(t0 )1−r + b(s) ds
1−r t0 1−r t0

or equivalently, setting H for the right-hand side of (1.28) and using that r < 1, we get
Z t
H(t)1−r ≤ a(t)1−r + (1 − r) b(s) ds.
t0

1
Since 1−r
> 0 and using (1.28)

1
 1

 Z t 
1−r
 Z t 
1−r
g(t) ≤ a(t)1−r + (1 − r) b(s) ds ≤ Cr a(t) + (1 − r) b(s) ds .
t0 t0

This concludes the proof of the lemma.

Remark 1.A.2. Call H the right-hand side of (1.28). If g is not continuous, note that the
function H is continuous and satisfies (1.28) (since b is positive and g ≤ H). Therefore,

76
 1.A. Some technical results

one can apply the lemma to H and then use the inequality g ≤ H.

Let us now state and prove the following result.

Lemma 1.A.3. Let b be a function such that limt→+∞ b(t) = 0. Pick a > 0 and γ ≥ 1.
Let g be a continuously differentiable positive function satisfying

γ+1
q
g 0 (t) ≤ −ag(t) 2 + b(t) g(t), t ≥ 0. (1.29)

Then, g(t) −→ 0.
t→+∞
γ
Proof. Pick ε > 0. Let t1 be a positive real such that for all t ≥ t1 , |b(t)| ≤ a2 ε 2 .
Step 1. We first show that there exists t∗ ≥ t1 , such that g(t∗ ) ≤ ε.
Assume, by way of contradiction, that it is not the case. Thus, one can consider the
√
function y = g, which satisfies

2y 0 (t) ≤ −ay(t)γ + b(t), t ≥ t1 . (1.30)

For all t ≥ t1 , we have

γ a γ a γ
2y 0 (t) ≤ −aε 2 + ε 2 ≤ − ε 2 .
2 2
As a consequence, for all t ≥ t1 ,
√ a γ
2 ε < 2y(t) ≤ 2y(t1 ) − (t − t1 ) ε 2 −→ −∞.
2 t→+∞

This is a contradiction.
Step 2. We show that for all t ≥ t∗ , g(t) ≤ ε.
Define T = inf{t ≥ t∗ , g(t) > ε}. By continuity of the function g, we have g(T ) = ε.
Hence,
γ+1 a γ√ a γ+1
g 0 (T ) ≤ −aε 2 + ε 2 ε < − ε 2 < 0.
2 2
Therefore, there exists δ0 > 0, such that for all 0 < δ ≤ δ0 ,

g(T + δ) < g(T ) = ε.

This is a contradiction with the definition of T .

77
Chapter 1 – Behavior of a time-inhomogeneous Kolmogorov type diffusion

Acknowledgements
The authors would like to thank Jürgen Angst for valuable exchanges and suggestions
about this work and Thomas Cavallazzi for his careful reading of the manuscript. We would
also like to thank the anonymous Referees for her/his careful reading of the manuscript
and useful advice.

78
 Chapter 2

B EHAVIOR OF A TIME - INHOMOGENEOUS

KINETIC L ÉVY - DRIVEN MODEL

Abstract: We study a one-dimensional kinetic stochastic model driven by a Lévy process with
a non-linear time-inhomogeneous drift. More precisely, the process (V, X) is considered, where
X is the position of the particle and its velocity V is the solution to a stochastic differential
equation with a drift of the form t−β F (v). The driving process can be a stable Lévy process of
index α or a general Lévy process under appropriate assumptions. The function F satisfies a
homogeneity condition and β is a real number. The behavior in large time of the process (V, X)
is proved and the precise rate of convergence is pointed out by using stochastic analysis tools.
To this end, we compute the moment estimates of the velocity process.

Keywords: kinetic stochastic equation; time-inhomogeneous stochastic differential equation;

Lévy process; explosion time; scaling transformation; asymptotic distribution; ergodicity; tight-
ness.

MSC2020 Subject Classification: Primary 60H10; Secondary 60F17; 60G52; 60G18; 60J65.

This chapter [GL21b] has been submitted.

79
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

2.1 Introduction
In this chapter, we consider a one-dimensional stochastic kinetic model driven by a
Lévy process.
Z t
dVt = dLt − F (Vt )t−β dt and Xt = X0 + Vs ds. (2.1)
0

The process (Vt , Xt )t>0 may be thought of as the velocity and position processes of a
particle subject to a friction force F (v)t−β and interacting with its environment.
Our purpose is to study the long-time behavior of solutions to (2.1) where L is an α-stable
(non-symmetric) Lévy process. More precisely, we look for the convergence in distribution
of the process (Vt/ε , Xt/ε )t>0 , as ε → 0, with an appropriate rate.
1 1
It is a simple observation when F = 0 to see that the rescaled process (ε α Vt/ε , ε1+ α Xt/ε )t>0
converges in distribution towards the Kolmogorov process (St , 0t Ss ds)t>0 , where S has
R

the same distribution as the driving process.

The goal of the present paper is to extend the results obtained in [GL21a], where the
driving process is a Brownian motion (α = 2).

The study of stochastic differential equations (SDEs) driven by a Lévy process is a

topic of great interest (see [Bas03] for a survey). The α-stable perturbation is a gener-
alization of the Gaussian case, and it is also motivated by some Langevin-type models
in stochastic climate dynamics (see [Dit99]). So far, most of the papers present results
about existence and uniqueness of solution, see for instance [AS09], [Don18], [Kur07],
[Pil13], [CSZ18] and [CZZ17]. The coefficients of the studied SDE are often supposed
to be time-homogeneous (see for instance [AS09] and [Don18]). Accordingly, the case of
time-dependent coefficients is scarcely studied (see [CZZ17], [Kur07] and [Zha13]). In this
situation, the usual tools associated with time-homogeneous equation may no longer be
invoked.
Furthermore, few papers (see [AS09], [Pri+12], [Rek20]) present results about the asymp-
totic behavior of the solution of such SDEs. For instance, in [AS09] the authors give
conditions for asymptotic stability of the solutions to a SDE driven by a Brownian mo-
tion and a compensated Poisson process, with coefficients that are supposed to satisfy
usual global Lipschitz and growth assumptions. In [Pri+12], the authors establish the
exponential ergodicity of the solutions to a SDE driven by an α-stable process, where the
drift coefficient is supposed to be the sum of two components, one linear and the other

80
 2.1. Introduction

bounded. In these papers, coefficients are time-homogeneous. In a number of articles,

the small noise influence of the solutions is analyzed. To our knowledge, the only works
considering the long-time behavior are [FT21], in a time-homogeneous setting, and the
present one.

Let us explain heuristically what the intuition of our analysis is. In long-time regime, we
observe three schemes, depending on the balance between the space and time coefficients of
the drift function with respect to α, the parameter of stability of the driving process. When
the drag force is sufficiently “small at infinity”, the convergence towards the Kolmogorov
process (S, 0· S) still holds. When the two terms in the stochastic equation of the velocity
R

process offset, we still get a kinetic process of the form (V, 0· V), as limiting process.
R

Though the process V no longer has the same distribution as the driving Lévy process.
Alternatively, when the drift swings with the random noise, the limiting process is no
longer kinetic.
Proofs are mainly based on moment estimates and on the self-similarity of the driving
process. By their scaling property, Lévy stable processes are natural extensions of the
Brownian motion. However, the jump component of the Lévy noise brings difficulties.
Indeed, by contrast with a Brownian motion, an α-stable Lévy process can only have
moments of order κ ∈ [0, α). Thus, moment estimation of the velocity process stands as a
significant part of our study (see Section 2.4). Moment estimates of Lévy and Lévy-type
processes were studied in [LP06], [Küh17] and [DS15]. Nevertheless, the methods used
can not be easily adapted to the solutions to a SDE. In fact, the key idea will be to
make a non-homogeneous cutting of the jumps size of the driving process. As explained
1
in [CM20] and references therein, the cutting threshold ξ 7→ ξ α makes appear integral
terms satisfying the scaling property (see the proof of Proposition 2.4.4).
The proof of the critical and sub-critical cases (see Theorem 2.2.4) significantly relies on
a change in both space and time, taking advantage of the scaling property of the driving
process, to be close to a stationary time-homogeneous SDE, as performed in [AW09] and
[GO13].
In addition, extensions for the solution to the SDE driven by a general Lévy process are
stated in Theorems 2.6.1 and 2.6.2. Let us point out that the case analyzed in Section 2.6.3
agrees with the equation (1) p. 1442 in [Dit99], where the described Langevin equation
has two noise terms, a white noise and a pure-jump noise. To this end, we first study the
asymptotic behavior of the rescaled Lévy driving process (rε Lt/ε )t≥0 , where rε is some

81
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

rate of convergence. Under appropriate assumptions on its Lévy measure, it converges in

distribution, as ε goes to zero (see Proposition 2.6.3 below).

Here is the structure of the paper. In Section 2.2, we introduce some notations and
state our main results. Theorems 2.2.2 and 2.2.4 are extension for the α-stable setting of
the main results stated in [GL21a] in case of a Brownian driving process. We study the
existence of the solution to (2.1) in Section 2.3. In Section 2.4, we give estimates of the
moment, which also ensure the non-explosion of the velocity process. The proofs of our
main results are presented in Section 2.5. Finally, Section 2.6 is devoted to the study of
the convergence of a rescaled Lévy process and to the extension of Theorem 2.2.2.

2.2 Notations and main results

Let (Lt )t≥0 be a Lévy process. Throughout the paper, we deal with L as an α-stable
Lévy process with α ∈ (0, 2). We call ν its Lévy measure, given by

a+ 1{z>0} + a− 1{z<0}
ν(dz) = dz, with a+ , a− ≥ 0 and a+ + a− > 0.
|z|1+α

As a Lévy measure, it satisfies R∗ (1∧z 2 )ν(dz) < +∞. By Lévy-Itô’s decomposition, L is a

R

pure-jump Lévy process and there exists a Poisson point measure N and its compensated
f such that, for all t ≥ 0,
Poisson measure N
Z t Z


 zN (ds, dz) if α ∈ (0, 1),
Z0 t ZR∗


 Z tZ
Lt = f(ds, dz) +
zN zN (ds, dz) if α = 1, (2.2)
Z0 t Z{0<|z|<1} 0 {|z|≥1}




if α ∈ (1, 2).


 z N (ds, dz)
f
0 R∗

In Section 2.6, the case of a generalized Lévy driving process will be discussed.
The space of continuous functions C((0, +∞), R) is endowed with the uniform topology

+∞
1
 
2
X
du : (f, g) ∈ C((0, +∞), R) →
7 n
min 1, sup |f − g| .
n=1 2 [ 1 ,n] n

Set

Λ := {λ : R+ → R+ , continuous and increasing function s.t. λ(0) = 0, lim λ(t) = +∞}

t→+∞

82
 2.2. Notations and main results

and 
1




1 if n
≤ t ≤ n,

kn (t) = n+1−t if n < t < n + 1,



0 if n + 1 ≤ t.


The space of right-continuous with left limits (càdlàg) functions D((0, +∞), R) is endowed
with the Skorokhod topology ds defined for (f, g) ∈ D((0, +∞), R)2 by
  
+∞
X 1   λ(t) − λ(s) 
n
1 ∧ inf a, ∃λ ∈ Λ, sup log ≤ a, sup |kn (t) (f ◦ λ(t) − g(t))| ≤ a .
n=1 2  s6=t t − s t≥ 1 
n

For simplicity, we shall write C and D for C((0, +∞), R) and D((0, +∞), R), respectively.
(ε)
For a family ((Zt )t>0 )ε>0 of càdlàg processes, we write

(ε)
(Zt )t>0 =⇒ (Zt )t>0 ,
ε→0

(ε)
if (Zt )t>0 converges in distribution to (Zt )t>0 in D, as ε → 0.
We write
(ε) f.d.d.
(Zt )t>0 =⇒ (Zt )t>0 ,
ε→0

(ε)
if for all finite subsets S ⊂ (0, +∞), the vector (Zt )t∈S converges in distribution to
(Zt )t∈S in RS , as ε → 0.
Let β a real number and F a continuous function satisfying

for some γ ∈ R, ∀v ∈ R, λ > 0, F (λv) = λγ F (v). (Hγ )

We introduce another assumption on F , which will sometimes be imposed in the sequel.

When (i) α ∈ (0, 1] or (ii) α ∈ (1, 2) and γ ≥ 1,

(Hsgn )
we suppose furthermore that for all v ∈ R, vF (v) ≥ 0.

We take an interest into the following one-dimensional stochastic kinetic model defined,
for t ≥ t0 > 0, by

dVt = dLt − t−β F (Vt ) dt, with Vt0 = v0 > 0, and dXt = Vt dt, with Xt0 = x0 ∈ R.
(SKEL )

83
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

In the following, sgn is the sign function with the convention that sgn(0) = 0. The ab-
breviation a.s. stands for almost surely. We denote by C some positive constants, which
may change from line to line. We use the subscripts to indicate the parameters on which
it depends. For instance, Ct0 ,α denotes a constant depending on the parameters t0 and α.

Remark 2.2.1. If a function π satisfies (Hγ ), then for all x ∈ R, π(x) = π(sgn(x)) |x|γ .

As an example of a function satisfying (Hγ ) one can keep in mind F : v 7→ sgn(v) |v|γ
(see also [GO13]).
Let us state our main results.

Theorem 2.2.2. Consider γ ∈ (1 − α2 , α). Assume that (Hγ ) and (Hsgn ) are satisfied,
and β > 1 + γ−1α
. Let (Vt , Xt )t≥t0 be the solution to (SKEL ) and (St )t≥0 be an α-stable
process, having same distribution as (Lt )t≥0 .
Then, in the space D,
 Z t 
1 1
1+ α
(ε Vt/ε , ε
α Xt/ε )t≥εt0 =⇒ St , Ss ds .
ε→0 0 t>0

Remark 2.2.3. Theorem 2.2.2 is also true when the following hypothesis holds instead
of (Hγ ).

F is such that (SKEL ) has a unique solution up to explosion and

(Hγ0 )
|F | ≤ G where G is a positive function satisfying (Hγ ).

For instance, the function F : v 7→ v

(1+v 2 )
(see also [FT21]) satisfies (Hγ0 ) (with γ = 0).

Theorem 2.2.4. Consider γ ∈ (1 − α2 , α) and β = 1 + γ−1 α

. Assume that (Hγ ) and (Hsgn )
are satisfied. Let (Vt , Xt )t≥t0 be the solution to (SKEL ).
Call Hf the eternal ergodic process solving the following SDE driven by an α-stable process

(Rt )t≥0 with same distribution as L and such that the distribution of H f
−∞ is the invariant

measure,
Hs  
dHs = dRs − ds − F Hs ds.
α
We denote by ΛF,t1 ,··· ,td the finite-dimensional distributions of H. f We call (V )
t t≥0 the pro-

cess having as finite-dimensional distributions the pushforward measure of ΛF,log(t1 ),··· ,log(td )
1/α 1/α
 1 
by the linear map T (u1 , · · · , ud ) := (t1 u1 , · · · , td ud ). Indeed, we have V = t α H
f
log(t) .
t≥0

84
 2.3. Existence up to explosion

Then, under (Hγ ), the following convergence, in the space D, holds

 Z t 
1 1
1+ α
(ε Vt/ε , ε
α Xt/ε )t≥εt0 =⇒ Vt , Vs ds .
ε→0 0 t>0

Theorem 2.2.5. Consider α > 1, γ ∈ [1, α) and β < 1 + γ−1 α

. Assume that (Hγ ) and
β
(Hsgn ) are satisfied. Let (Vt , Xt )t≥t0 be the solution to (SKEL ). Define q := α+γ−1 < α1 .
c the ergodic process solving the following SDE driven by an α-stable process (R )
Call H t t≥0

with same distribution as L and starting at its invariant measure,

dHs = dRs − F (Hs ) ds.

 
Call ΠF its invariant measure. We call (Vt )t≥0 the process whose f.d.d. are T ∗ Π⊗d F : the
pushforward measure of Π⊗d q q
F by the linear map T (u1 , · · · , ud ) := (t1 u1 , · · · , td ud ).
Then,
f.d.d.
 
εq Vt/ε =⇒ (Vt )t≥0 .
t≥εt0 ε→0

Remark 2.2.6. As we will see in Section 2.3, the assumption γ > 1 − α2 is needed in
order to obtain the existence up to explosion of the solution under the hypothesis (Hγ ).

Remark 2.2.7. Let us point out that, during the proof of Theorem 2.2.2 and Theo-
rem 2.2.4, we employ some moment estimates for the solution V . We state the estimates
below.
Assume that (Hsgn ) is satisfied. We suppose also that the hypothesis on the sign of F holds
for (α, γ, κ) ∈ (1, 2)×[0, 1]×(1, α). Then, for any α ∈ (0, 2), γ ∈ R, β ∈ R and κ ∈ [0, α),
there exists a constant Cγ,κ,β,t0 such that,

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 t α .

Note that the above bounds are the best possible, taking F = 0.

2.3 Existence up to explosion

In this section, we study the existence of the solution to (SKEL ) up to explosion time.

Remark 2.3.1. Assume that (Hγ ) holds. If 0 < γ < 1, then the function F is γ-Hölder
and if γ ≥ 1, it is locally Lipschitz.

85
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Proposition 2.3.2. Assume that (Hγ ) is satisfied. There exists a pathwise unique strong
solution to (SKEL ), defined up to the explosion time, provided that
α
(i) 1 − 2
< γ < 1 and β ≥ 0 when α ∈ (0, 2).
(ii) γ ≥ 1 when α > 1.

Proof. If γ ∈ (0, 1), the drift coefficient is γ-Hölder (see Remark 2.3.1) and locally
bounded, thereby the conclusion of the first point follows from Remark 1.3 in [CZZ17].
Assume now that α > 1 and γ ≥ 1. The drift coefficient is locally Lipschitz (see Re-
mark 2.3.1) and locally bounded, so we can apply Lemma 115 p. 78 in [Sit05] to get the
pathwise uniqueness. Thanks to Theorem 137 p. 104 in [Sit05], it suffices to prove that
there exists a weak solution.
The drift coefficient is continuous with respect to its two variables, so it is a locally
bounded and measurable function. By a standard localization argument, using Theorem
9.1 p. 231 in [IW81], since the drift coefficient is locally Lipschitz, there is a unique solution
defined up to explosion.

2.4 Moment estimates and non-explosion of the ve-

locity process
In this section, we present estimates on moments of the velocity process V solution to
(SKEL ). This will be useful to conclude of the non-explosion of solution to (SKEL ) with
Lemma 2.4.1, and to control some terms appearing along the proofs of Theorem 2.2.2 and
2.2.4 in Section 2.5.
Let V be the unique solution up to explosion time to (SKEL ). For all r ≥ 0, define the
stopping time
τr := inf{t ≥ t0 , |Vt | ≥ r}. (2.3)

Set τ∞ := limr→+∞ τr the explosion time of V .

We give first a sufficient condition for the non-explosion of a general process.

Lemma 2.4.1. Let (Yt )t≥t0 be a càdlàg process and τ∞ its explosion time. Assume that
there exist two measurable and non-negative functions φ and b such that
(i) φ is non-decreasing and limr→∞ φ(r) = +∞,
(ii) b is finite-valued,

86
 2.4. Moment estimates and non-explosion of the velocity process

(iii) and for all t ≥ t0 ,

sup E [φ(|Yt∧τr |)] ≤ b(t). (2.4)
r≥0

Then τ∞ = +∞ a.s.

Proof. Pick t ≥ t0 . Using the definition of τr , the monotony of φ and (iii), we get, for all
r ≥ 0,
h i
φ(r)P(τr ≤ t) ≤ E φ (Yτr ) 1{τr ≤t} ≤ E [φ (Yτr )] ≤ b(t).

Thus, by Fatou’s lemma,

1
0 ≤ P (τ∞ ≤ t) ≤ lim inf P(τr ≤ t) ≤ b(t) lim = 0.
r→∞ r→∞ φ(r)

As a consequence,
X
0 ≤ P (τ∞ < +∞) ≤ P (τ∞ ≤ t) = 0.
t∈Q

This concludes the proof.

We will show that there exists a constant Cγ,κ,β,t0 such that

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 t α . (2.5)

Proposition 2.4.2. Pick α ∈ (0, 1). Assume that (Hsgn ) holds. Recall that (Vt )t≥t0 is
the solution to (SKEL ). For any γ, β, the explosion time τ∞ is a.s. infinite and for all
κ ∈ [0, α), there exists a constant Cκ,t0 such that, we have

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t α . (2.6)

Proof. Fix t ≥ t0 . Since α < 1, the stable process can be written as

Z tZ X
Lt = zN (ds, dz) = ∆Ls .
0 R∗ s≤t

q
Fix κ ∈ [0, α). Pick the sequence of C 2 -functions fn : x 7→ x2 + n1 , which converges
uniformly to the function x 7→ |x| on R. Then, for all n ≥ 1, we apply Itô’s formula (see

87
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Theorem 32 p. 78 in [Pro05]) to get

Z t∧τr
fn0 (Vs )F (Vs ) Z t∧τr Z
fn (Vt∧τr ) = fn (v0 ) − ds + (fn (Vs− + z) − fn (Vs− )) N (ds, dz)
t0 sβ t0 R∗
X
≤ fn (v0 ) + (fn (Vs− + ∆Ls ) − fn (Vs− )).
s≤t∧τr

The term tt∧τ fn0 (Vs )F (Vs )s−β ds is non-negative, since (Hsgn ) holds.
R r
0

Hence, the previous equation can be written as

X
fn (Vt∧τr ) ≤ fn (v0 ) + (fn (Vs ) − fn (Vs− )).
s≤t∧τr

Since kfn0 k∞ ≤ 1, we deduce that (fn (Vs ) − fn (Vs− )) ≤ |∆Vs | = |∆Ls |, hence,
X
|Vt∧τr | ≤ fn (Vt∧τr ) ≤ fn (v0 ) + |∆Ls | .
s≤t∧τr

Furthermore, since κ < α < 1, we have

 κ

|Vt∧τr |κ ≤ fn (v0 )κ + 
X
|∆Ls | .
s≤t∧τr

Taking the expectation, we get

 κ 

E [|Vt∧τr |κ ] ≤ E [fn (v0 )κ ] + E 

X
|∆Ls |  .
s≤t

Notice that the process L+ |∆Ls | is an α-stable process. Then, since κ < α,
P
t := s≤t
letting n → +∞, we obtain
h κi κ
E [|Vt∧τr |κ ] ≤ |v0 |κ + E L+
t ≤ Ct0 ,κ t α .

Thanks to Lemma 2.4.1, we can conclude that the explosion time of V is a.s. infinite, and
(2.6) follows, letting r → ∞.

Proposition 2.4.3. Pick α ∈ (1, 2). Recall that (Vt )t≥t0 is the solution to (SKEL ). For
any γ ∈ [0, 1) and any β ∈ R, the explosion time τ∞ is a.s. infinite and for all κ ∈ [0, 1],

88
 2.4. Moment estimates and non-explosion of the velocity process

there exists Cγ,κ,β,t0 such that we have


κ γ−1
κ
t α

if α
+ 1 ≤ β,
∀t ≥ t0 , E [|Vt | ] ≤ Cγ,κ,β,t0 (2.7)
tκ 1−β
1−γ else.


Proof of Proposition 2.4.3. Assume that γ ∈ [0, 1) and fix κ ∈ [0, 1]. Then Jensen’s in-
equality yields, for all t ≥ t0 , E [|Vt |κ ] ≤ E [|Vt |]κ , hence it suffices to verify (2.7) only for
κ = 1.
Recall that under (Hγ ), there exists a positive constant K, such that for all v ∈ R,
|F (v)| ≤ K |v|γ . Hence, we can write, for any t ≥ t0 and r ≥ 0,
Z t∧τr
V(t∧τr )− ≤ |v0 − Lt0 | + L(t∧τr )− + s−β |F (Vs∧τr )| ds
t0
Z t∧τr
≤ |v0 − Lt0 | + L(t∧τr )− + K s−β |Vs∧τr |γ ds.
t0

Since L is an α-stable process, it has a finite first moment, which can be computed. Taking
the expectation in the above inequality, we get, by choosing Ct0 big enough,
h i h i Z t
E V(t∧τr )− ≤ E [|v0 − Lt0 |] + E L(t∧τr )− + K s−β E [|Vs∧τr |γ ] ds
t0
Z t
1 γ
≤ Ct0 t + K
α s−β E [|Vs∧τr |] ds.
t0

h i
Recalling that τr is given by (2.3), the function gr : t 7→ E V(t∧τr )− is bounded by r.
Applying a Grönwall-type lemma (see Lemma 2.A.1), we end up, for β 6= 1, with
 ! 1 
h i 1−γ1
1−γ
∀t ≥ t0 , E V(t∧τr )− ≤ Cγ Ct0 t + αK(t1−β − t01−β ) .
1−β

The case β = 1 can be treated similarly. Thanks to Lemma 2.4.1, we conclude that the
explosion time of V is a.s. infinite, and (2.7) follows from Fatou’s lemma.

Proposition 2.4.4. Pick α ∈ [1, 2). Assume here that for all v ∈ R, vF (v) ≥ 0. Recall
that (Vt )t≥t0 is the solution to (SKEL ). For any γ ∈ R and any β ∈ R, the explosion time
τ∞ is a.s. infinite and there exists Cκ,t0 such that

κ
for κ ∈ (0, α), ∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t α . (2.8)

89
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Proof. The key idea is to slice the small and big jumps in a non-homogeneous way with
1
respect to the function ξ 7→ ξ α . We write the proof in the general setting of α ∈ (1, 2).
When α = 1, the proof is similar since ν is symmetric.
Pick ξ ≥ t0 . As explained in [CM20] and references therein, by using this cutting threshold,
the α-stable Lévy driving process can be written as
Z tZ Z tZ Z tZ
Lt − Lt0 = 1
f(ds, dz) +
zN 1 zN (ds, dz) − 1 zν(dz) ds.
t0 |z|≤ξ α t0 |z|>ξ α t0 |z|>ξ α

The two first integrals then satisfy the same scaling property as the α-stable Lévy driving
process.
We can compute Z
a+ − a− 1 −1
1 zν(dz) = ξα .
|z|>ξ α α−1
Step 1. We first apply Itô’s formula and estimate the expectation of each term for κ ≤ 1,
in order to get (2.8).
Fix η > 0 and define the C 2 -function f : v 7→ (η + v 2 )κ/2 . For all t ≥ t0 , by Itô’s formula,
using that for all v ∈ R, vF (v) ≥ 0, we have

a+ − a− 1 −1 Z t
f (Vt∧τr ) ≤ f (V0 ) − ξα 1{s≤τr } f 0 (Vs ) ds + Mt + Rt + St , (2.9)
α−1 t0

where Z tZ
Mt := 1 1{s≤τr } [f (Vs− + z) − f (Vs− )] N
f(ds, dz), (2.10)
t0 0<|z|<ξ α

Z tZ
Rt := 1 1{s≤τr } [f (Vs− + z) − f (Vs− )])N (ds, dz), (2.11)
t0 |z|≥ξ α

Z tZ
St := 1 1{s≤τr } [f (Vs + z) − f (Vs ) − zf 0 (Vs )] ν(dz) ds. (2.12)
t0 0<|z|<ξ α

Note that, since κ < 1, for all v ∈ R,

κ−1
|f 0 (v)| ≤ κη 2 . (2.13)

Moreover, remark that for all k > α,

Z
a+ + a− k −1
1 |z|k ν(dz) = ξα , (2.14)
0<|z|<ξ α k−α

90
 2.4. Moment estimates and non-explosion of the velocity process

and for all k < α, Z

a+ + a− k −1
1 |z|k ν(dz) = ξα . (2.15)
|z|≥ξ α α−k
We estimate expectations of M , R and S.
To that end, we first show that the local martingale (Mt )t≥t0 is a martingale.

Fix q ≥ 2 and r ≥ 0. Set

Z tZ
It (q) := 1 1{s≤τr } |f (Vs− + z) − f (Vs− )|q ν(dz) ds.
t0 0<|z|<ξ α

1
Notice that, since for all |v| ≤ r and |z| ≤ ξ α , |f (v + z) − f (v)| ≤ f 0 1[−(r+ξ α1 ),r+ξ α1 ] |z|,
∞
so we have
q Z tZ
It (q) ≤ f 0 1[−(r+ξ α1 ),r+ξ α1 ] 1 1{s≤τr } |z|q ν(dz) ds.
∞ t0 0<|z|<ξ α

Hence, it is a finite quantity, since q ≥ 2 and (2.14) holds. Therefore, for q ≥ 2, by

Kunita’s inequality (see Theorem 4.4.23 p. 265 in [App09]), there exists Dq > 0 such that
" #
q
 h q i 
E sup |Ms | ≤ Dq E It (2) 2 + E [It (q)] < +∞.
t0 ≤s≤t

Hence, by Theorem 51 p. 38 in [Pro05], M is a martingale.

We estimate now the finite variation part S defined in (2.12). We use a similar idea as in
the proof of Theorem 3.1 p. 3863 in [DS15]. Note that for all v ∈ R,
κ κ
|f 00 (v)| = κ(2 − κ)v 2 (v 2 + η) 2 −2 + κ(v 2 + η) 2 −1
κ κ
= κ(2 − κ)v 2 (v 2 + η)−1 (v 2 + η) 2 −1 + κ(v 2 + η) 2 −1
κ
≤ κ(3 − κ)(v 2 + η) 2 −1
κ
≤ κ(3 − κ)η 2 −1 , since κ
2
− 1 < 0.

1
Assume that |z| < ξ α . Using Taylor’s formula, we get a.s.

1 κ
|f (Vs + z) − f (Vs ) − zf 0 (Vs )| ≤ κ(3 − κ)η 2 −1 z 2 .
2

91
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Hence, we get the almost sure following bound

Z
0 1 κ
−1
Z
2
1 (f (Vs + z) − f (Vs ) − zf (Vs )) ν(dz) ≤ κ(3 − κ)η 2
1 z ν(dz).
0<|z|<ξ α 2 0<|z|<ξ α

Injecting (2.14), we get

Z
1 κ a+ + a− 2 −1
1 (f (Vs + z) − f (Vs ) − zf 0 (Vs )) ν(dz) ≤ κ(3 − κ)η 2 −1 ξα . (2.16)
0<|z|<ξ α 2 2−α

It remains to study the Poisson integral R defined in (2.11), using Theorem 2.3.7 p. 106
in [App09]. Pick κ ≤ 1, by Hölder property of power functions, we can write,
 κ  κ  κ  κ
|f (v + z) − f (v)| = η + (v + z)2 2
− (v + z)2 2
+ (v + z)κ − v κ + v 2 2
− η + v2 2

κ
≤ 2η 2 + |z|κ .

We deduce that
Z Z
κ 1
1 |f (Vs + z) − f (Vs )| ν(dz) ≤ η 2 ν(|z| ≥ ξ α ) + 1 |z|κ ν(dz).
|z|≥ξ α |z|≥ξ α

Injecting (2.15), this leads to

Z
κ a+ + a− −1 a+ + a− κ −1
1 |f (Vs + z) − f (Vs )| ν(dz) ≤ η 2 ξ + ξα . (2.17)
|z|≥ξ α α α−κ

Gathering (2.13), (2.17) and (2.16), we get

E [|Vt∧τr |κ ] ≤ E [f (Vt∧τr )] ≤ E [f (Vt0 )] + tξ −1 ×

κ−1 a+ − a−
κ/2 a+ + a− a+ + a− κ 1 −1 a+ + a− α
 
1 κ 2
κη 2 ξ +η
α + ξ + κ(3 − κ)η
α 2 ξ . (2.18)
α−1 α α−κ 2 2−α
2
Choosing η = t α and ξ = t, we get

a+ − a− a+ + a− a+ + a− 1 a+ + a−
 
κ
E [|Vt∧τr |κ ] ≤ E [f (Vt0 )] + t α × κ + + + κ(3 − κ)
α−1 α α−κ 2 2−α
κ
≤ Cκ,t0 t α .
(2.19)
Thanks to Lemma 2.4.1, we can conclude that the explosion time of V is a.s. infinite and

92
 2.4. Moment estimates and non-explosion of the velocity process

letting r → +∞, for all κ ∈ [0, 1],

κ
E [|Vt |κ ] ≤ Cκ,t0 t α . (2.20)

Step 2. Pick κ ∈ (1, α). We estimate R in another way, using again Theorem 2.3.7 p. 106
in [App09].
By Hölder property of power function and (2.15), we get
Z Z κ

1 |f (Vs + z) − f (Vs )| ν(dz) ≤ 1 2zVs + z 2 2

ν(dz)
|z|≥ξ α |z|≥ξ α
! (2.21)
a+ + a− κ −1 κ a+ + a− κ
−1
≤ Cκ ξ α + |Vs | 2 κ ξ
2α .
α−κ α− 2

Gathering (2.16), (2.21) and then using that for all v ∈ R, |f 0 (v)| ≤ κ |v|κ−1 ,

a+ + a− κ −1 1 a+ + a− 2 −1
 
κ κ
E [|Vt∧τr | ] ≤ E [f (Vt0 )] + Cκ ξ α + κ(3 − κ)η 2 −1 ξα t
α−κ 2 2−α
a+ − a− 1 −1 Z t h κ−1 i a+ + a− κ −1 Z t h κi
+κ ξα E |Vs | ds + Cκ ξ 2α E |Vs | 2 ds. (2.22)
α−1 t0 α − κ2 t0

2
Injecting (2.20), and choosing η = t α and ξ = t, we get

κ
E [|Vt∧τr |κ ] ≤ Cκ,t0 ,α t α .

Taking r → +∞, (2.8) follows.

Example 2.4.5. Remark that the velocity process V is more explicit in the linear case
(γ = 1), and that the moment estimate is as best as possible. Choose F (1) = ρ > 0,
F (−1) = −ρ. Pick β 6= 1, so
! !
t1−β Z t s1−β
Vt = v0 + exp −ρ exp ρ dLs
1 − β t0 1−β

is solution to (SKEL ).
Hence, by an integration by parts,
Z t
1 1−β t1−β s1−β
−t1−β )
Vt = v0 + Lt − eρ 1−β (t0 Lt0 − e−ρ 1−β ρs−β eρ 1−β Ls ds.
t0

93
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Thus,
1 1 1
 
E [|Vt |] ≤ Ct0 t α + t1−β+ α ≤ Ct0 t α .

The case β = 1 can be treated similarly.

2.5 Proof of the asymptotic behavior of the solution

This section is devoted to the proofs of our main results, Theorems 2.2.2, 2.2.4 and
2.2.5.
Notice that, in the super-critical end critical regimes, it suffices to prove the convergence
1
of the rescaled velocity process (ε α Vt/ε )t≥εt0 in the space D endowed with the Skorokhod
topology. Assume for a moment that this convergence is proved.
For ε ∈ (0, 1] and t ≥ εt0 we can write
Z t
1 1
1+ α 1+ α
ε Xt/ε = ε x0 + Vs(ε) ds.
εt0

 
Let us introduce the mapping gε : V 7→ Vt , εtt 0 Vs ds defined and valued on D. Clearly,
R
t>0
(ε)
the theorem will be proved once we show that gε (V• ) converges weakly in D endowed
with the Skorokhod topology. This mapping is converging, as ε → 0, to the continuous
 
mapping g : V 7→ Vt , 0t Vs ds
R
.
t>0
(ε) 1
To see V (ε) as a process of D([0, +∞)), we state, for all s ∈ [0, εt0 ], Vs(ε) := Vεt0 = ε α v0 .
Call Pε , P the distribution of V (ε) , S, respectively. Invoking the Portmanteau theorem
(see Theorem 2.1 p. 16 in [Bil99]), it suffices to prove that for all bounded and uniformly
continuous function h : D([0, +∞)) × D([0, +∞)) → R,
Z Z
h(gε (ω)) dPε (dω) −→ h(g(ω)) dP (dω).
D([0,+∞))2 ε→0 D([0,+∞))2

Pick such a function h. By assumption, the convergence Pε =⇒ P holds, hence, using

ε→0
Lemma 2.A.4, it suffices to prove that the uniformly bounded sequence (h ◦ gε ) of con-
tinuous functions on D([0, +∞)) converges to the continuous function h ◦ g uniformly on
compact subsets of D([0, +∞)). Let K be a compact subset of D([0, +∞)). Then, for
all ω ∈ K, max[0,εt0 ] |ω| is uniformly bounded by a constant, say M . Fix η > 0. By the
uniform continuity of h, there exists δ > 0 such that for all ω ∈ K,

du (gε (ω), g(ω)) ≤ δ =⇒ |h ◦ gε (ω) − h ◦ g(ω)| ≤ η.

94
 2.5. Proof of the asymptotic behavior of the solution

There exists ε1 > 0 small enough, such that for all ε ≤ ε1 , for all ω ∈ K,
Z εt0
du (gε (ω), g(ω)) ≤ C ω(s) ds ≤ Cεt0 M ≤ δ.
0

Therefore, we proved that it suffices to prove the convergence of the rescaled velocity
1
process (ε α Vt/ε )t≥εt0 in order to prove Theorems 2.2.2 and 2.2.4.
In Sections 2.5.1 and 2.5.2, the aim is to prove the convergence of the velocity process.

2.5.1 Asymptotic behavior in the super-critical regime

γ−1
In the remainder of this section, we assume that γ ≥ 0 and β > 1 + α
.

Proof of Theorem 2.2.2. Thanks to a change of variables, we have, for all ε ∈ (0, 1] and
t ≥ εt0 ,
Z t/ε
1 1 1 1
ε α Vt/ε =ε α (v0 − Lt0 ) + ε α Lt/ε − ε α F (Vs )s−β ds
t0
Z t
1 1 1
=ε (v0 − Lt0 ) + ε Lt/ε − ε
α α
β−1+ α
F (Vu/ε )u−β du.
εt0
1
By self-similarity, L(ε) := (ε α Lt/ε )t≥0 has the same distribution as an α-stable process.
As a consequence, thanks to Theorem 3.1 p. 27 in [Bil99] and Lemma 2.A.3, it suffices to
prove
(ε) (ε) P
∀T > 0 sup Vt − Lt −→ 0, as ε → 0. (2.23)
εt0 ≤t≤T

Recall that under (Hγ ), there exists a positive constant K, such that
!
γ V•(ε) γ
ε α F 1 ≤ K V•(ε) . (2.24)
ε α

Modifying the factor in front of the integral, we get

!
(ε) 1 (ε) β−1+ 1−γ
Z t
γ Vu(ε)
Vt = ε (v0 − Lt0 ) +
α Lt −ε α ε F
α
1 u−β du. (2.25)
εt0 εα

95
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Gathering (2.25) and (2.24), for all T > 0, we have,

!
(ε) (ε) 1
β−1+ 1−γ
Z t
γ Vu(ε)
sup Vt − Lt ≤ε (v0 − Lt0 ) + ε
α α sup ε F
α
1 u−β du
εt0 ≤t≤T εt0 ≤t≤T εt0 ε α
Z T
1
β−1+ 1−γ γ
≤ε (v0 − Lt0 ) + ε
α α K Vu(ε) u−β du.
εt0

Taking the expectation and using the moment estimates on V (see Remark 2.2.7), we
obtain, when β 6= αγ + 1,
"Z #
(1−γ) T γ
Z T
γi
β−1+ 1−γ
h
−β
ε β−1+ α
E K Vu(ε) u du = ε α KE Vu(ε) u−β du
εt0 εt0
Z T
1
h γi
=ε β−1+ α
KE Vu/ε u−β du
εt0
Z T
1−γ γ
≤ εβ−1+ α Cα,β,t0 u α −β du
εt0
γ
 
β−1+ 1−γ γ
−β+1 −β+1 1
= Cα,β,t0 ε α T α − t0 α
ε α .

1−γ 1 γ−1
Hence, setting r := min(β − 1 + α
, α) which is positive, since β > 1 + α
, we get
" #
(ε) (ε)
E sup Vt − Lt = O (εr ).
εt0 ≤t≤T ε→0

γ
The case β = 1 + α
can be treated similarly to get
" #
(ε) (ε) 1
E sup Vt − Lt = O (ε α ln(ε)).
εt0 ≤t≤T ε→0

This concludes the proof.

Remark 2.5.1. Observe that, in this proof, we did not use the condition “γ < 1 or for
all v ∈ R, vF (v) ≥ 0”, except to get moment estimates.

2.5.2 Asymptotic behavior in the critical regime

We adapt the Proposition 2.1, p. 187 of [GO13] to the α-stable Lévy case.
Pick a C 2 -diffeomorphism ϕ : [0, t1 ) → [t0 , +∞). Let V be the solution to the equation
(SKEL ). Thanks to Proposition 3.4.1 p. 124 in [ST94], the following process is also an

96
 2.5. Proof of the asymptotic behavior of the solution

α-stable process !
Z t
dLϕ(s)
(Rt )t≥0 := 1 . (2.26)
0 ϕ0 (s) α t≥0

Then, by the change of variables t = ϕ(s), we get

Z t
1
Z t
F (Vϕ(s) ) 0
Vϕ(t) − Vϕ(0) = ϕ0 (s) α dRs − ϕ (s) ds.
0 0 ϕ(s)β

Thanks to an integration by parts, we get

1
ϕ0 (s)1− α ϕ00 (s) Vϕ(s)
!
Vϕ(s)
d 1 = dRs − F (Vϕ(s) ) ds − ds.
ϕ0 (s) α ϕ(s)β αϕ0 (s) ϕ0 (s) α1

Set Ω = D([t0 , ∞)) the set of càdlàg functions, that equal ∞ after their (possibly infinite)
explosion time. Introduce the scaling transformation Φϕ defined, for ω ∈ Ω, by

ω(ϕ(s))
Φϕ (ω)(s) := 1 , with s ∈ [0, t1 ).
ϕ0 (s) α

As a consequence, we obtain the following result.

Proposition 2.5.2. If V is a solution to the equation (SKEL ), then V (ϕ) := Φϕ (V ) is a

solution to
1
ϕ0 (s)1− α 0 1 ϕ00 (s) Vs(ϕ) (ϕ) Vϕ(0)
dVs(ϕ) = dRs − F (ϕ (s) αV
s
(ϕ)
) ds − ds, with V0 = 1 , (2.27)
ϕ(s)β ϕ0 (s) α ϕ0 (0) α

where R is given by (2.26).

Conversely, if V (ϕ) is a solution to (2.27), then Φ−1ϕ (V
(ϕ)
) is a solution to the equation
(SKEL ), where Z t
1
Lt − Lt0 := (ϕ0 ◦ ϕ−1 ) α (s) dRϕ−1 (s) .
t0

Furthermore, uniqueness in law, pathwise uniqueness, strong existence hold for the equa-
tion (SKEL ) if and only if they hold for the equation (2.27).

In the following, we will focus on the exponential change of time ϕe : t 7→ t0 et . This

scaling is convenient since it allows to produce a time-homogeneous term in (2.27). Thanks
to Proposition 2.5.2, the process V (e) := Φe (V ) satisfies the SDE driven by an α-stable

97
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

process (Rt )t≥0 ,

Vs(e) 1− 1 −β 1
 1 s 
dVs(e) = dRs − ds − t0 α e(1− α −β)s F t0α e α Vs(e) ds. (2.28)
α
γ−1
Proof of Theorem 2.2.4. Assume in the sequel that β = 1 + α
.

Step 1. Firstly we prove the finite-dimensional convergence of the rescaled velocity pro-
cess. To that end, we reduce the problem to the convergence of a time-homogenous process.
Since (Hγ ) holds, (2.28) becomes

Vs(e)  
dVs(e) = dRs − ds − F Vs(e) ds. (2.29)
α

Using the bijection Φe induced by the exponential change of time (see Proposition 2.5.2),
and the unique strong existence of the velocity process V (see Proposition 2.3.2 and
Remark 2.2.7), there exists a pathwise unique strong solution H to the time-homogeneous
equation (2.29). Hence, we have the equality
!
Vt0 et
= (Ht )t≥0 ,
(t0 et )1/α t≥0

as two solutions to the same SDE, starting from the same point. We can write the above
equality as
Vt
 
1 = (Hlog(t/t0 ) )t≥t0 .
t α t≥t0
So, we have, for all ε > 0, d ∈ N∗ , and (t1 , · · · , td ) ∈ [εt0 , +∞)d ,
!
Vε−1 t1 Vε−1 t  
−1 1/α
, · · · , −1 d1/α = Hlog(t1 )+log((εt0 )−1 ) , · · · , Hlog(td )+log((εt0 )−1 ) . (2.30)
(ε t1 ) (ε td )

Since lim sup|x|→+∞ −F (x)−x/α

x
< 0, it follows from Proposition 0.1 in [Kul09] that the
process (Ht )t≥0 is exponentially ergodic. We denote its invariant measure by ΛF . Call
f the solution to the time homogeneous equation (2.29), such that the initial condition
H
d
H −∞ has the distribution ΛF . For t1 , · · · , td ∈ R , let ΛF,t1 ,··· ,td := L(Ht1 , · · · , Htd ) be
f f f
f ,··· ,H
the distribution of (H f ). Then, for all s ≥ 0, Λ
t1 td F,t1 ,··· ,td = ΛF,t1 +s,··· ,td +s . Indeed,

thanks to the invariance property of ΛF , (H• ) and (H•+s ) satisfy the same SDE, starting
f f

98
 2.5. Proof of the asymptotic behavior of the solution

from the same point. As a consequence, we get the stationary limit

 
lim L H
f
log(t1 )+log((εt0 )−1 ) , · · · , Hlog(td )+log((εt0 )−1 ) = ΛF,log(t1 ),··· ,log(td ) .
f (2.31)
ε→0

Moreover, by exponential ergodicity, we have for every continuous and bounded function
ψ : Rd → R,
h  i h  i
E ψ Hlog(t1 /(t0 ε)) , · · · , Hlog(td /(t0 ε)) −E ψ H
f
log(t1 /(t0 ε)) , · · · , Hlog(td /(t0 ε))
f −→ 0.
ε→0
(2.32)
We postpone the proof of this limit in Step 2.
To conclude this step, gather (2.30), (2.31) and (2.32) to get
!
Vε−1 t1 Vε−1 td
, · · · , =⇒ ΛF,log(t1 ),··· ,log(td ) .
(ε−1 t1 ) /α
1
(ε−1 td )1/α ε→0

This can also be written as

1 1
 
ε α Vt1 /ε , · · · , ε α Vtd /ε =⇒ T ∗ ΛF,log(t1 ),··· ,log(td ) ,
ε→0

where T ∗ ΛF,log(t1 ),··· ,log(td ) is the pushforward of the measure ΛF,log(t1 ),··· ,log(td ) by the linear
1/α 1/α
map T (u1 , · · · , ud ) := (t1 u1 , · · · , td ud ).

Step 2. Let us now prove (2.32).

For the sake of clarity, let us give a proof for d = 2, the general case d ≥ 2 is similar.
−1
Let ψ : R2 → R be a continuous and bounded function. Pick εt0 ≤ s ≤ t. Set h0 = v0 t0 α ,
(2.32) is now equivalent to
       
E ψ Hlog(s/(t0 ε)) , Hlog(t/(t0 ε)) H0 = h0 − E ψ Hlog(s/(t0 ε)) , Hlog(t/(t0 ε)) H0 ∼ ΛF −→ 0.
ε→0

 
We set µε := L Hlog(s/(t0 ε)) H0 = h0 . We now use the generalized Markov property of
solution to SDE driven by Lévy process. For the sake of completeness, we state and prove
it in our context in Appendix (see Lemma 2.A.6). This leads to
       
E ψ Hlog(s/(t0 ε)) , Hlog(t/(t0 ε)) H0 = h0 = E ψ H0 , Hlog(t/s) H0 ∼ µε

99
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

and, since ΛF is invariant,

       
E ψ Hlog(s/(t0 ε)) , Hlog(t/(t0 ε)) H0 ∼ ΛF = E ψ H0 , Hlog(t/s) H0 ∼ ΛF .

Then, we are reduced to prove

       
E ψ H0 , Hlog(t/s) H0 ∼ µε − E ψ H0 , Hlog(t/s) H0 ∼ ΛF −→ 0.
ε→0

The left-hand side can be written as,

Z    
E ψ H0 , Hlog(t/s) H0 = y (µε (dy) − ΛF (dy)) .
R

Hence, setting p(t, x, dy) := Px (Ht ∈ dy) and [Link] V for the total variation norm, we get
       
E ψ H0 , Hlog(t/s) H0 ∼ µε − E ψ H0 , Hlog(t/s) H0 ∼ ΛF
Z
≤ kψk∞ |p (log(s/(t0 ε)), h0 , dy) − ΛF (dy)|
R

≤ kψk∞ kp (log(s/(t0 ε)), h0 , ·) − ΛF kT V .

This converges to 0, as ε → 0, by the exponential ergodicity of H.

Step 3. Let us prove now the tightness of the family of distributions of the càdlàg process
   1 
V (ε) = ε α Vt/ε on every compact interval [m, M ], 0 < m ≤ M .
t≥εt0 t≥εt0
We check the Aldous’s tightness criterion stated at Theorem 16.10 p.178 in [Bil99]. Let
a, η, T be positive reals. Let τ be a discrete stopping time with finite range T , bounded
by T . Choose δ > 0 and ε > 0 small enough.
We have, by Jensen’s inequality, for r = α2 ,
"Z #r
h ri h ri τ +δ γ
(ε) (ε) −β
E Vτ +δ − Vτ(ε) ≤E Lτ +δ − L(ε)
τ +E K Vu(ε) u du .
τ

Since L(ε) is an α-stable process, by the strong Markov property,

h
(ε) ri r
E Lτ +δ − L(ε)
τ = E [ELτ [|Lδ − L0 |r ]] ≤ Cδ α .

100
 2.5. Proof of the asymptotic behavior of the solution

The stopping time has a finite range T . Hence, we can write

"Z # " "Z ##
τ +δ γ τ +δ γ
E K Vu(ε) u−β du = E E K Vu(ε) u−β du τ
τ τ
" " # #
1 Z τi +δ
γ
K Vu(ε) u−β du 1{τ =τi }
X
=E E 1{τ =τi }
τi ∈τ P(τ = τi ) τi
" "Z # #
X 1 τi +δ
(ε) γ −β
≤E E K Vu u du 1{τ =τi } .
τi ∈τ P(τ = τi ) τi

(γ−1)
For each τi ∈ T , using the relation β = 1 + α
and the moment estimates on V (see
Remark 2.2.7), we obtain
"Z #
τi +δ γ
Z τi +δ h γi
−β
E K Vu(ε) u du = KE Vu(ε) u−β du
τi τi
Z τi +δ  1

γ 1
−β
≤K u α du = K (τi + δ) − τi
α α

τi
1
≤ Kδ 1, α .

1 1 1
The term δ 1, α has to be read as δ or δ α depending on the fact that x 7→ x α is a Lipschitz
continuous function on [0, T + δ], if α < 1, or a α1 -Hölder function, if α > 1.
By Markov’s inequality, for δ small enough, we have
r

(ε) Kδ r, α

P Vτ +δ − Vτ(ε) ≥a ≤ ≤ η.
ar

Furthermore, by moment estimates (see Propositions 2.4.2, 2.4.3 and 2.4.4), for all t ≥ εt0 ,

(ε) r r
h i
sup Vt ≤ Ct α .
ε

Hence, using again Markov’s inequality, by Corollary and Theorem 16.8 p. 175 in [Bil99],
this concludes the proof of the tightness of the velocity process and therefore the proof of
Theorem 2.2.4.

101
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

2.5.3 Asymptotic behavior in the sub-critical regime

γ−1
Assume in this section that β < 1 + α
and α > 1. As a consequence, αq < 1. We
  1

take an interest into the power change of time ϕq : t 7→ + (1 − αq)t t1−αq

. Thanks0
1−αq

(q)
to Proposition 2.5.2, the process V := Φq (V ) satisfies the SDE driven by an α-stable
process R distributed as L,
 
dVs(q) = dRs − F Vs(q) ds − qϕqαq−1 Vs(q) ds. (2.33)

For simplicity, we shall write φ instead of φq .

Proof of Theorem 2.2.5. Step 1. We first prove the finite dimensional convergence of the
(ε)
velocity process (Vt )t≥εt0 := (εq Vt/ε )t≥εt0 . We give a proof for d = 2, the general case
d ≥ 2 is similar.
We call H the ergodic process solution to
 
dHs = dRs − F Hs ds, with H0 = h0 := v0 t−q
0 . (2.34)

We denote by ΠF its invariant measure. Using the bijection induced by the power change
of time (Proposition 2.5.2), as solutions to the same SDE starting at the same point, we
have, for all ε > 0, and (s, t) ∈ [εt0 , +∞)2 ,

Vε−1 s Vε−1 t
   
(q) (q)
εq q , εq q = Vϕ−1 (ε−1 s) , Vϕ−1 (ε−1 t)
s t

Using Theorem 3.1 p. 27 in [Bil99], it suffices to prove that for all (s, t) ∈ [εt0 , +∞)2
   
(q) (q)
• Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) − Vϕ−1 (ε−1 s) , Vϕ−1 (ε−1 t) −→ 0, where k·k is a metric
ε→0
on R2 .
 
• Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) =⇒ ΠF ⊗ ΠF .
ε→0
(q) κ
h i
Step 2. Pick κ ∈ (1, α). We prove that E Ht − Vt −→ 0.
t→+∞
We have
   
(q) (q)
d H − V (q) = − F (Ht ) − F (Vt ) dt + qϕαq−1 (t)Vt dt.
t

102
 2.5. Proof of the asymptotic behavior of the solution

By straightforward differentiation, we can write

κ (q) (q) κ−1

d H − V (q) = −κ F (Ht ) − F (Vt ) Ht − Vt dt
t
(q) κ−1
 
(q) (q)
+ κqϕαq−1 (t)Vt sgn Ht − Vt Ht − Vt dt. (2.35)

We set
(q) κ
h i
g(t) := E Ht − Vt , t ≥ 0.

Taking expectation in (2.35), we get

 
(q) (q) κ−1
g 0 (t) = −κE F (Ht ) − F (Vt ) Ht − Vt
 
(q) κ−1
 
(q) (q)
+ κqϕαq−1 (t)E Vt sgn Ht − Vt Ht − Vt .

Since γ ≥ 1, the function F −1 is γ1 -Hölder, therefore there exists Cγ > 0 such that,
   
0 (q) κ−1+γ αq−1 (q) (q) κ−1
g (t) ≤ −Cγ E Ht − Vt + κ |q| ϕ (t)E Vt Ht − Vt .

Then, by Jensen’s inequality, since γ ≥ 1,

 
κ−1+γ (q) (q) κ−1
g 0 (t) ≤ −Cγ g(t) κ + κ |q| ϕαq−1 (t)E Vt Ht − Vt .

Using Hölder’s inequality and moment estimates (Proposition 2.4.2), we have

κ−1+γ 1 κ−1
g 0 (t) ≤ −Cγ g(t) κ + C |q| ϕ(αq−1)(1− α ) (t)g(t) κ , g(0) = 0.

1
Note that since α > 1 and αq < 1, then ϕ(αq−1)(1− α ) (t) −→ 0, therefore the conclusion
t→+∞
follows from Lemma 2.A.7.
h
(q) κi
Besides, for all t ≥ εt0 , E Hϕ−1 (ε−1 t) − Vϕ−1 (ε−1 t) = g (ϕ−1 (ε−1 t)) −→ 0.
ε→0
2
Step 3. Pick (s, t) ∈ [εt0 , +∞) . Similarly, as in [GL21a], one can prove that the solution
H to (2.34) satisfies
 
Hϕ−1 (ε−1 s) , Hϕ−1 (ε−1 t) =⇒ ΠF ⊗ ΠF . (2.36)
ε→0

103
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

2.6 Extended results in the Lévy case

In this section, the driving process of (SKEL ) is supposed to be a general Lévy process
L. We denote by (A, ν, b) its generating triplet, with respect to the truncation function
h : z 7→ −1 ∨ (z ∧ 1). Here ν is the Lévy measure, A > 0 and b ∈ R. This means that, for
t ≥ 0, in virtue of the Lévy-Khintchine formula (see Theorem 8.1 p. 37 in [Sat99]), the
characteristic function of Lt is given by

h i Au2 Z
E eiuLt = etψ(u) , where ψ(u) := − + ibu + (eiux − 1 − iuh(x))ν(dx).
2 R

Pick α ∈ (0, 2). The Lévy measure ν of the driving process is supposed to satisfy either

g(z)
ν(z) = 1{z6=0} , where g is a non-negative measurable function such that
|z|1+α (H1ν,α )
c+ := lim g(z) ≥ 0, c− := lim g(z) ≥ 0,
z→+∞ z→−∞

or Z
for some α0 > 1, |z|α0 ν(dz) < +∞. (H2ν,α0 )
|z|≥1

Remark that if ν satisfies (H1ν,α ) with α > 1, then it satisfies (H2ν,α0 ). Note that any tem-
pered stable process satisfies (H1ν,α ), and any truncated α-stable process satisfies (H2ν,α0 ).
For clarity, we recall the stochastic kinetic model: for t ≥ t0 > 0,

dVt = dLt − t−β F (Vt ) dt, with Vt0 = v0 > 0, and dXt = Vt dt, with Xt0 = x0 ∈ R.
(SKEL )
We work under the assumption that there exists a unique solution to (SKEL ). We will
show that Theorem 2.2.2 can be extended with this general Lévy driving process. We
suppose first the Lévy process to be without a Brownian component. This case will be
discussed in Section 2.6.3. We obtain the two following theorems, depending on which
hypothesis is satisfied by the Lévy measure ν.

Theorem 2.6.1. Consider γ ∈ [0, α) and β ≥ 0. Assume that (H1ν,α ) and (Hsgn ) are

104
 2.6. Extended results in the Lévy case

satisfied and define


γ



 α
if g is bounded,


γ
if α ∈ (0, 1),



α



γ
if α ∈ (1, 2), γ ∈ [0, 1) and b = 0,


α
pα (γ) :=
γ if α ∈ [1, 2), γ ∈ [0, 1) and b 6= 0,







γ if α ∈ (1, 2), γ = 1,







γ

+ γ
if α ∈ (1, 2), γ ∈ (1, α).
α 2

Let (Vt , Xt )t≥t0 be the solution to (SKEL ).

Then there exist a rate of convergence εθ and a Lévy process L, given in Proposition 2.6.3
(iii)-(v), such that, as ε → 0, if β > 1 + pα (γ) − θ, in the space D,
 Z t 
(εθ Vt/ε , ε1+θ Xt/ε )t≥εt0 =⇒ Lt , Ls ds .
ε→0 0 t>0

Theorem 2.6.2. Consider γ ≥ 0 and β ≥ 0. Assume that (H2ν,α0 ) is satisfied and define





γ if γ ∈ [0, 1),

γ
pα0 (γ) :=
min(γ, α0 + γ2 ) if γ ∈ [0, 1) and vF (v) ≥ 0,


γ +γ

if γ ∈ [1, α0 ] and vF (v) ≥ 0.
α0 2

Suppose that β > 1 + pα0 (γ) − 1. Let (Vt , Xt )t≥t0 be the solution to (SKEL ).
Then there exists a Lévy process L, given in Proposition 2.6.3 (iii), such that, as ε → 0,
in the space D,  Z  t
(εVt/ε , ε2 Xt/ε )t≥εt0 =⇒ Lt , Ls ds .
ε→0 0 t>0

2.6.1 Large time behavior of the Lévy driving process

Since the Lévy noise is no longer self-similar, we need to study its large-time behavior.
We dedicate this subsection to the study of the convergence in distribution of the rescaled
(ε)
noise (Lt )t≥0 := (rε Lt/ε )t≥0 , for a suitable rate rε , tending to 0.

105
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

We introduce another assumption on ν, which will sometimes be imposed in the sequel:

g(z)
ν(z) = 1{z6=0} , where g is a non negative measurable function with
|z|1+α
(Hg )
Z +∞
|g(z) − g(−z)|
dx < +∞.
1 zα

Inspired from [RT11], we get the following result.

Proposition 2.6.3. Let (Lt )t≥0 be a Lévy process having generating triplet (A, ν, b), with
respect to the truncation function h : z 7→ −1 ∨ (z ∧ 1).
(i) Assume that the Lévy measure satisfies the condition (H1ν,α ) with α = 1, c+ = c− = c
 
and (Hg ). Then the process εLt/ε converges in distribution to the 1-stable Lévy
t≥0
∗ ∗
process L generated by (0, ν , b ), where
Z +∞
g(z) − g(−z) c1{z6=0}
b∗ := b + (z − h(z)) dz and ν ∗ (dz) := dz.
1 z 1+α |z|2

(ii) Suppose that the Lévy measure satisfies the condition (H1ν,α ) with 1 < α < 2 and
(Hg ). If
∗
Z +∞
g(z) − g(−z)
b := b + (z − h(z)) dz 6= 0,
1 z 1+α
 
then the process εLt/ε converges in distribution to the deterministic Lévy process
t≥0
L generated by (0, 0, b∗ ).
(iii) If the Lévy measure satisfies (H2ν,α0 ) and b∗ := b + |z|≥1 (z − h(z)) ν(dz) 6= 0, then
R
 
the process εLt/ε converges in distribution to the deterministic Lévy process L∗
t≥0
generated by (0, 0, b∗ ).

(iv) Assuming that 0 < α < 1, if the Lévy measure satisfies the condition (H1ν,α ), then
 1 
the process ε α Lt/ε converges in distribution to the α-stable Lévy process L with
t≥0

c+ 1{z>0} + c− 1{z<0}
ν ∗ (dz) := dz.
|z|1+α

(v) Assuming that 1 < α < 2 and b = 0, if the Lévy measure satisfies the condition
 1 
(H1ν,α ), then the process ε α Lt/ε converges in distribution to the Lévy process L
t≥0

106
 2.6. Extended results in the Lévy case

with measure and center

c+ 1{z>0} + c− 1{z<0 } Z
ν ∗ (dz) := dz, b∗ := zν ∗ (dz).
|z|1+α R∗

Proof. By the Lévy-Khintchine’s formula, the generating triplet of (L(ε) )t≥0 := (rε Lt/ε )t≥0
is given by
r2
Aε = ε A, (2.37)
ε

for all B ∈ B(R), ν ε (B) = ε−1 ν({z, zrε ∈ B}), (2.38)

rε Z    
ε
b = b+ rε−1 h(rε z) − h(z) ν(dz) . (2.39)
ε R∗

Call (A∗ , ν ∗ , b∗ ) the generating triplet of the limiting process L. By Corollary 3.6 p. 415
in [JS03] and Theorem 14.7 p. 81 in [Sat99], we have to check that

bε −→ b∗ , (2.40)
ε→0

Z Z
∗
ε
A + 2
h (z)ν (dz) ε
−→ A + h2 (z)ν ∗ (dz), (2.41)
R∗ ε→0 R∗

and that for any continuous and bounded function f which is zero in a neighborhood of
zero, Z Z
f (z)ν ε (dz) −→ f (z)ν ∗ (dz). (2.42)
R∗ ε→0 R∗

(i) Recall that α = 1 and assume that the Lévy measure ν satisfies the conditions
(H1ν,α ) and (Hg ).
To prove (2.40), we write bε as

ε
Z +∞
g(z) − g(−z)  −1 
b =b+ ε h(εz) − h(z) dz.
0 z 1+α

The dominated convergence theorem can be applied, and we show that it converges
to
∗
Z +∞
g(x) − g(−x)
b =b+ (x − h(x)) dx.
1 x2
Observe that the condition (Hg ) was only required for this step.
Afterwards, note that, using a change of variables,

107
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Z Z
g(zrε−1 )
h2 (z)ν ε (dz) = rεα ε−1 h2 (z) dz,
R∗ R∗ |z|1+α

and thus, we can apply the dominated convergence theorem to prove that the last
integral converges to Z
h2 (z)ν ∗ (dz).
R∗

Let f be a continuous and bounded function which is zero in a neighborhood of zero,

then, using again a change of variables and applying the dominated convergence
theorem, Z Z
ε
f (z)ν (dz) −→ f (z)ν ∗ (dz).
R∗ ε→0 R∗

(ii) The proof is similar to the previous one and thus is left to the reader.
(iii) In this point, we assume that (H2ν,α0 ) holds.
The convergence (2.40) follows from the dominated convergence theorem.
Using the explicit form of the truncation function, we get, for rε < 1,
Z Z
h2 (z)ν ε (dz) = ε−1 h2 (rε z)ν(dz)
R∗ RZ∗ Z Z
=ε 2
z ν(dz) + 2
εz ν(dz) + ε−1 ν(dz).
0<|z|<1 1≤|z|≤ε−1 |x|>ε−1
(2.43)
Using the property of a Lévy measure, the first term in (2.43) converges to zero, as
ε → 0, by assumption. Then, the last two terms in (2.43) are lower than
Z Z
εα0 −1 |z|α0 ν(dz) + εα0 −1 |z|α0 ν(dz),
1<|z|<ε−1 |z|≥ε−1

which converges to zero, when ε goes to zero, since α0 > 1.

Let f be a continuous and bounded function and assume that there exists δ > 0
such that f (z) = 0 for all |z| ≤ δ, thus,
Z Z Z
ε
f (z)ν (dz) = −1
ε f (rε z)ν(dz) ≤ Cε α0 −1
|z|α0 ν(dz).
R∗ |z|> ηε |z|> ηε

This vanishes as ε → 0.
(iv) Take 0 < α < 1.

108
 2.6. Extended results in the Lévy case

Using the explicit form of h, giving in Proposition 2.6.3, we have,

Z −ε− α1 Z +∞
1 1 1
−1 −1 −1
ε
b =ε α b+ [ε α − ε ]ν(dz) + 1 [ε−1 − ε α −1 ]ν(dz)
−∞ ε− α
Z −1 Z ε− α1
1 1
−1 −1
+ε α
1 (z + 1)ν(dz) + ε α (z − 1)ν(dz).
−ε− α 1

Since g has finite limits at infinity, for any δ > 0, we can choose η > 1 big enough
so that |g(z) − c+ | < δ for z ≥ η, and |g(z) − c− | < δ for z ≤ −η.
Hence,

Z ε− α1 Z ε− α1
1 1
−1 −1
lim sup ε α (z − 1)ν(dz) = lim sup ε α (z − 1)ν(dz)
ε→0 1 ε→0 η
Z ε− α1
1
−1 + x−1 c+ + δ
≤ lim sup ε α (c + δ) dx = .
ε→0 η x1+α 1−α

Similarly,

Z ε− α1 Z ε− α1
1 1
−1 −1
lim inf ε α (z − 1)ν(dz) = lim inf ε α (z − 1)ν(dz)
ε→0 1 ε→0 η
Z ε− α1
1
−1 + z−1 c+ − δ
≥ lim inf ε α (c − δ) dz = .
ε→0 η z 1+α 1−α

The choice of δ being arbitrary, we get

Z ε− α1
1
−1 c+
lim ε α (z − 1)ν(dz) = .
ε→0 1 1−α

If ε is small enough, then we can upper bound

Z +∞
1 1 1
lim sup 1 [ε−1 − ε α −1 ]ν(dz) ≤ lim sup(ε−1 − ε α −1 )(c+ + δ)ν([ε− α , +∞))
ε→0 ε− α ε→0
1 c+ + δ c+ + δ
= lim sup(1 − ε α ) = .
ε→0 α α

Moreover,
Z +∞
1 1 1
lim inf 1 [ε−1 − ε α −1 ]ν(dz) ≥ lim inf (ε−1 − ε α −1 )(c+ − δ)ν([ε− α , +∞])
ε→0 ε− α ε→0
1 c+ − δ c+ − δ
= lim inf (1 − ε α ) = .
ε→0 α α

109
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Similarly, this leads to

1
−1
Z −1
c−
lim ε α
1 (z − 1)ν(dz) = − ,
ε→0 −ε− α 1−α

and,
Z −ε− α1
1 c−
lim [ε α −1 − ε−1 ]ν(dz) = − .
ε→0 −∞ α
Hence, we obtain
c+ − c−
lim bε = .
ε→0 α(1 − α)
c −c + −
Since, R∗ h(z)ν ∗ (dz) = α(1−α)
R
, the drift coefficient of the limiting process equals
zero.
The proof of (2.41) and (2.42) are identical to the one done in (i).
(v) Take 1 < α < 2 and assume that b = 0.
After a change of variables, we can apply the dominated convergence theorem to
1
ε
Z  1 1
−α
 g(yε− α )
b = h(y) − ε h(yεα ) dy.
R∗ |y|1+α

The proof of (2.41) and (2.42) are identical to the one done in (i).

2.6.2 Proofs of Theorems 2.6.1 and 2.6.2

In this section, we suppose that β > 1 + pα (γ) − θ, where θ is the exponent of the rate
of convergence given in Proposition 2.6.3 and pα (γ) is given in the statement of Theorems
2.6.1 and 2.6.2. Recall that (Vt )t≥t0 is the solution to (SKEL ).

Moment estimates of the velocity process

As in Section 2.4, we will show that there exists a constant Cγ,κ,β,t0 such that

∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 ,b tpα (γ,κ) , (2.44)

where pα (γ, κ) has to be detailed.

110
 2.6. Extended results in the Lévy case

Proposition 2.6.4. Pick α ∈ (0, 1). Assume (H1ν,α ) and (Hsgn ). For any γ, β, the ex-
plosion time τ∞ is a.s. infinite and for all κ ∈ [0, α), there exists Cκ,t0 ,b such that, we
have
κ κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 ,b t α , or equivalently pα (γ, κ) = .
α
Proof. The proof is analogous to the proof of Proposition 2.4.2. The term tt0 bfn0 (Vs ) ds
R

is bounded by |b| t since kfn0 k∞ ≤ 1. Moreover, since ν satisfies (H1ν,α ) with α < 1, the
process L+ s≤t |∆Ls | satisfies the conditions of Theorem 3.1 c) in [DS15]. Thus, for
P
t :=
all κ ∈ [0, α), for all t ≥ t0 ,
h κi κ
E L+ t ≤ Ct0 ,κ t α .

The estimates for V follows.

Proposition 2.6.5. Assume either (H2ν,α0 ) or (H1ν,α ) with α ∈ (1, 2). For any γ ∈ [0, 1)
and any β ∈ R the explosion time τ∞ is a.s. infinite and for all κ ∈ [0, 1], there exists
Cγ,κ,β,t0 and Cγ,t0 such that under (H1ν,α ), we have

κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 t α + Cγ,t0 |b|κ tκ . (2.45)

Or equivalently pα (γ, κ) = ακ if b = 0, and pα (γ, κ) = κ, else.

And under (H2ν,α0 ), there exists Cγ,κ,β,t0 ,b such that we have

∀t ≥ t0 , E [|Vt |κ ] ≤ Cγ,κ,β,t0 ,b tκ .

Proof. We explain the differences with respect to the proof of Proposition 2.4.3. Under
each hypothesis on the Lévy measure ν, the Lévy process has a finite first moment. We
write Lt as the sum bt + L
b , where L
t
b is the Lévy process without the drift part.

We get,
h i h i Z t
E V(t∧τr )− ≤ Ct0 |b| t + E L
b
(t∧τr )− +K s−β E [|Vs∧τr |]γ ds.
t0

The proof of Theorem 3.1 (a) p. 3861 in [DS15] can be adapted to estimate the moment
of the Lévy process stopped at the stopping time τr , given by (2.3).
(i) If the Lévy measure ν satisfies (H1ν,α ) with 1 < α < 2, then it satisfies the conditions
of Theorem 3.1 (a) and (c) in [DS15]. Thus, for all κ ∈ [0, 1], there exists Ct0 ,κ such

111
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

that, for all t ≥ t0 ,

h κi
E L
b
(t∧τr )− ≤ Ct0 ,κ t,

and,
h κi κ
E L
b
t ≤ Ct0 ,κ t α .

(ii) If the Lévy measure ν satisfies (H2ν,α0 ), then it satisfies the conditions of Theorem
3.1 (a) in [DS15]. Thus, for all κ ∈ [0, 1], there exists Ct0 ,κ such that, for all t ≥ t0 ,
h κi
E L
b
(t∧τr )− ≤ Ct0 ,κ t.

Applying the Grönwall-type lemma again (see Lemma 2.A.1) and Fatou’s lemma, for
β 6= 1, we end up with
 ! 1 
h i 1−γ 1−γ
∀t ≥ t0 , E V(t∧τr )− ≤ Cγ Ct0 ,b t + K(t1−β − t01−β ) .
1−β

The case β = 1 can be done in a similar manner.

We conclude that the explosion time of V is a.s. infinite. To refine the estimates under
the hypothesis (H1ν,α ), we apply again the Grönwall-type lemma to
h i Z t
E [|Vt |] ≤ Ct0 |b| t + E L
b
t +K s−β E [|Vs |]γ ds.
t0

γ−1
This proves (2.45) since α
+ 1 − β ≤ 0.

Proposition 2.6.6. Assume that for all v ∈ R, vF (v) ≥ 0. Pick γ ∈ R and β ∈ R. For
each of the following cases, the explosion time τ∞ is a.s. infinite.
(i) Assume that (H1ν,α ) holds with a bounded function g and α ∈ [1, 2), there exists Cκ,t0
such that

κ κ
for κ ∈ (0, α), ∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t α , or equivalently pα (γ, κ) = . (2.46)
α

(ii) Assume that (H1ν,α ) holds with α ∈ [1, 2). Then,

for κ ∈ [0, 1] (resp. κ ∈ [0, 1), if α = 1), there exists Cκ,t0 such that

∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 tκ , or equivalently pα (γ, κ) = κ; (2.47)

112
 2.6. Extended results in the Lévy case

for κ ∈ (1, α), γ ∈ [0, 1) and b = 0, there exists Cκ,t0 such that

3κ 3κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t 2α , or equivalently pα (γ, κ) = ; (2.48)
2α

for κ ∈ (1, α), there exists Cκ,t0 such that

κ κ κ κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t α + 2 , or equivalently pα (γ, κ) = + . (2.49)
α 2

(iii) Assume (H2ν,α0 ). Then,

for κ ∈ [0, 1], there exists Cκ,t0 such that,

∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 tκ , or equivalently p(γ, κ) = κ; (2.50)

for κ ∈ [0, α0 ], there exists Cκ,t0 such that

κ
+κ κ κ
∀t ≥ t0 , E [|Vt |κ ] ≤ Cκ,t0 t α0 2
, or equivalently p(γ, κ) = + . (2.51)
α0 2

Proof. We highlight only the differences with respect to the proof of Proposition 2.4.4. In
the following, we assume that α 6= 1, the proof is similar for α = 1.

Step A. Assume that (H1ν,α ) holds with a bounded function g.

Step A1. Pick κ ∈ [0, 1]. We adapt the estimates of the Itô’s formula’s terms.
There is an additional term in (2.9), given, for t ≥ t0 , by
Z t
1{s≤τr } f 0 (Vs )b ds.
t0

If g is a bounded function, the other terms of (2.9) can be estimated in the same way,
and (2.19) becomes

κ 1 κ
E [|Vt∧τr |κ ] ≤ E [f (Vt0 )] + t α +1− α κ |b| + t α sup |g| ×
R
a+ − a− a+ + a− a+ + a− 1 a+ + a−
 
κ
κ + + + κ(3 − κ) ≤ Cκ,t0 ,b t α . (2.52)
α−1 α α−κ 2 2−α

This gives the proof of (2.46) for κ ∈ [0, 1].

113
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Step A2. Pick κ ∈ (1, α). We estimate R, given by (2.11) in another way.
The inequality (2.22) becomes

a+ + a− κ −1 1 a+ + a− 2 −1
 
κ κ
E [|Vt∧τr | ] ≤ E [f (Vt0 )] + t Cκ ξ α + κ(3 − κ)η 2 −1 ξα
α−κ 2 2−α
a+ − a− 1 −1
Z t h
a+ + a− κ −1 Z t h
 i κi
κ−1
+κ ξ α + |b| E |Vs | ds + Cκ ξ 2α E |Vs | 2 ds. (2.53)
α−1 t0 α − κ2 t0

The inequality (2.46) follows as in the proof of Proposition 2.4.4.

Step B. When working under (H2ν,α0 ), we will pick κ ≤ α0 during the proof. Assume
that either (H1ν,α ) holds with an unbounded function g or (H2ν,α0 ) holds, then we are not
able to estimate the following terms
Z Z Z
1 zν(dz), 1 |z|k ν(dz), and 1 |z|k ν(dz).
|z|>ξ α 0<|z|<ξ α |z|≥ξ α

Hence, we apply the same proof scheme with small and big jumps sliced at 1.

Step B1. This leads to a similar bound as in (2.19):

Z t
E [|Vt∧τr |κ ] ≤ E [f (Vt∧τr )] ≤ E [f (Vt0 )] + |b| E [|f 0 (Vs∧τr )|] ds + E [t ∧ τr ] ×
t0
!
Z
1
κ κ
Z
η κ/2
ν(|z| ≥ 1) + |z| ν(dz) + κ(3 − κ)η 2 −1 z 2 ν(dz) ≤ Cκ,t0 t.
|z|≥1 2 0<|z|<1

By Jensen’s inequality, we can deduce (2.47) and (2.50).

Step B2. Pick κ ∈ (1, α). We estimate R, given by (2.11) in another way.
By classical Hölder inequality,
Z Z κ Z
κ
|f (Vs + z) − f (Vs )| ν(dz) ≤ 2zVs + z 2 2
ν(dz) ≤ C(1 + |Vs | 2 ) |z|κ ν(dz)
|z|≥1 |z|≥1 |z|≥1
(2.54)
The last integral is finite. Gathering (2.16), (2.54) and then using (2.47), (2.50) or Propo-
sition 2.6.5,
Z t Z t
h κ i h i
κ
E [|Vt∧τr | ] ≤ E [f (Vt∧τr )] ≤ Cκ,t0 t + C E |Vs | 2 ds + |b| E |Vs |κ−1 ds.
t0 t0

114
 2.6. Extended results in the Lévy case

Taking r → +∞, we can conclude that pα (γ, κ) = 1 + pα (γ, κ/2).

Step B3. We refine the estimates.

Fix κ ∈ [0, α). There exists ε1 such that, κ ≤ α − ε and α − ε > 1. Hence, we can write
h i
E |Vt |α−ε ≤ Cκ,t0 t1+pα (γ,(α−ε)/2) .

Using Jensen’s inequality, we get

κ
E [|Vt |κ ] ≤ Cκ,t0 t α−ε (1+pα (γ,(α−ε)/2)) ,

and it suffices to let ε → 0 to conclude. This concludes the proof of (2.48) and (2.49)
The last step is identical under (H1ν,α ) and can be done with α0 instead of α − ε, under
(H2ν,α0 ). This concludes the proof of (2.51).

Proofs of Theorems 2.6.1 and 2.6.2

We are now in position to give the proofs of Theorems 2.6.1 and 2.6.2. Assume that
either (H1ν,α ) or (H2ν,α0 ) is satisfied.
As in the α-stable case (see Section 2.5), it suffices to prove the convergence of the rescaled
(ε)
velocity process (Vt )t>0 := (rε Vt/ε )t>0 .
Thanks to Proposition 2.6.3, there exists θ ∈ {1, α1 } such that L(ε) := (εθ Lt/ε )t≥0 converges
in distribution.
For the sake of simplicity, we omit the dependencies of p with respect to α and α0 . Assume
that γ ≥ 0 and β > 1 + p(γ) − θ.
We can write, for T > 0,
Z T
(ε) (ε) γ
sup Vt − Lt ≤ rε (v0 − Lt0 ) + rε1−γ εβ−1 K Vu(ε) u−β du.
εt0 ≤t≤T εt0

Using the moment estimates on V (see Section 2.6.2), this leads, with β 6= p(γ) + 1, to
"Z #
T γ  
−β p(γ)−β+1
rε1−γ εβ−1 E K Vu(ε) u du ≤ C rε εβ−1−p(γ) T p(γ)−β+1 − t0 rε .
εt0

115
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

Hence, setting q := min(β−1+θ−p(γ), θ), which is a positive number since β > 1+p(γ)−θ,
we obtain " #
(ε) (ε)
E sup Vt − Lt = O(εq ).
εt0 ≤t≤T ε→0

The case β = 1 + p(γ) can be treated similarly. This concludes both of the proofs.

2.6.3 Lévy driving process with a Brownian component

We extend the results given in Theorems 2.2.2, 2.2.4, 2.6.1 and 2.6.2 to a driving
process having a Brownian component. To this end, we decompose the Lévy noise as the
sum of a Brownian part (Bt )t≥0 and a Lévy part without a Brownian component (Lt )t≥0 .
Remark 2.6.7. The reasoning is the same with a Lévy process L b instead of the Brownian

driving process. Indeed, let rbε be the rate such that (rbε L
b )
t/ε t≥0 converges. The conclusion
−1
follows, provided that rbε rε tends to 0.
Pick ρ ≥ 0. We consider the following one-dimensional stochastic kinetic model, for
t ≥ t0 > 0,

(2) (2) (2) γ (2)

dVt = dLt + dBt − t−β ρ sgn(Vt ) Vt dt, Vt0 = v0 ∈ R,
(2) (2) (2)
(SKE2)
and dXt = Vt dt, Xt0 = x0 ∈ R.

We compare this solution with the solution to the equation driven by a Lévy process L
without a Brownian component. From now on, for the sake of convenience, the solutions to
(1) (1)
(SKEL ) with F (v) = ρ sgn(v) |v|γ , will be denoted by (Vt , Xt )t≥t0 instead of (Vt , Xt )t≥t0 .
The asymptotic behavior of the latter SDE is given by Theorems 2.2.2, 2.2.4, 2.6.1 and
2.6.2. We will show that the Brownian part has no contribution.
Let us first point out some results about existence up to explosion of (SKE2).
Proposition 2.6.8.
(i) When γ ≥ 1, (SKE2) admits a unique solution up to explosion.
(ii) Pick α ∈ (1, 2). If (SKE2) is driven by the sum of a Brownian motion and an
α-stable process, then, if 0 ≤ γ < 1, there exists a solution to (SKE2).
Proof.
(i) Since the drift coefficient is locally Lipschitz, by a standard localization argument,
using Theorem 3.1 and Remark 2 p. 338-339 in [Kun04], there is a unique solution
defined up to explosion.

116
 2.6. Extended results in the Lévy case

(ii) Since the drift coefficient is a continuous function, using a standard localization
argument, we can apply Theorem 3.1 p. 866 in [Kur07] to conclude.

Assume in the following that (SKE2) admits a unique solution up to explosion.

(2) (2)
Theorem 2.6.9. Consider γ ∈ [0, α) and β ≥ 0. Let (Vt , Xt )t≥t0 be a solution to
(SKE2). The Lévy noise is supposed to be the sum of a Brownian noise (Bt )t≥0 and a
Lévy process (Lt )t≥0 without Brownian component. Suppose also that (SKEL ) satisfies the
conditions of either Theorems 2.2.2, 2.2.4, 2.6.1 or 2.6.2. Then (V (2) , X (2) ) has the same
asymptotic behavior as (V (1) , X (1) ), which is given in the cited theorems.

Proof. For i ∈ {1, 2}, let us introduce the following rescaled processes:

(i,ε) (i) (i,ε) (i) (ε) √

(Vt )t≥εt0 := (rε Vt/ε )t≥εt0 , (Xt )t≥εt0 := (εrε Xt/ε )t≥εt0 , and (Bt )t≥0 := ( εBt/ε )t≥0 .

We write, for all t ≥ εt0 ,

! !!
(1,ε) (2,ε) − 12 (ε)
Z t
Vu(1,ε) Vu(2,ε)
Vt − Vt = rε Bt0 − rε ε Bt − rε1−γ εβ−1 rεγ F −F u−β du.
εt0 rεγ rεγ

Hence, by Itô’s formula, we get

Z t 
(2,ε) 2 1
  
(1,ε)
Vt − Vt = rε2 Bt20 − 2 Vs(1,ε) − Vs(2,ε) rε ε− 2 dBs(ε) + rε2 ε−1 (t − εt0 )
εt0
! !!
Z t   Vu(1,ε) Vu(2,ε)
− 2rε1−γ εβ−1 Vs(1,ε) − Vs(2,ε) rεγ F −F u−β du.
εt0 rεγ rεγ

The last term on the right-hand side of the upper equality is positive since F is an
increasing function.
Moreover, we can apply BDG inequalities (see Theorem 73 p. 222 in [Pro05]) to the local
martingale, defined by
Z t 
1

Mt := −2 Vs(1,ε) − Vs(2,ε) rε ε− 2 dBs(ε) , t ≥ εt0 .
εt0

This yields, for each T > 0,

" # " Z T  #
2
E sup Mt2 ≤E 4rε2 ε−1 Vs(1,ε) − Vs(2,ε) ds .
εt0 ≤t≤T εt0

117
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

For the sake of clarity, given x and y, we introduce x . y to mean that there exists a
positive constant C such that x ≤ Cy. We get
" # " #
(2,ε) 4
  h i
(1,ε)
E sup Vt − Vt . rε4 E Bt40 + rε4 ε−2 T 2 +E sup Mt2
εt0 ≤t≤T εt0 ≤t≤T
h i Z T  2 
. rε4 E Bt40 + rε4 ε−2 T 2 + 4rε2 ε−1 E Vs(1,ε) − Vs(2,ε) ds.
εt0

Hence, using that for all real a, a2 ≤ a4 + 1, we deduce that

" #
(2,ε) 4
 
(1,ε)
E sup Vt − Vt
εt0 ≤t≤T
Z T " #
(2,ε) 4
h i  
(1,ε)
. rε4 E Bt40 + rε4 ε−2 T 2 + 4rε2 ε−1 T + 4rε2 ε−1 E sup Vt − Vt ds.
εt0 εt0 ≤t≤s

Applying Grönwall’s lemma, up to stopping times, we get

" #
(2,ε) 4
   h i   
(1,ε)
E sup Vt − Vt . rε4 E Bt40 + rε4 ε−2 T 2 + 4rε2 ε−1 T exp 4rε2 ε−1 T .
εt0 ≤t≤T

We deduce also that

" # " #
(2,ε) 4 (2,ε) 4
   
(1,ε) (1,ε)
E sup Xt − Xt ≤ T 4E sup Vt − Vt
εt0 ≤t≤T εt0 ≤t≤T
 h i   
. T 4 rε4 E Bt40 + rε4 ε−2 T 2 + 4rε2 ε−1 T exp 4rε2 ε−1 T .
 
(1,ε) (2,ε) (1,ε) (2,ε)
Hence, Vt − Vt , Xt − Xt converges uniformly towards 0 in probability
t≥εt0
on compacts, as ε goes to zero.
The conclusion follows from Theorems 2.2.2, 2.2.4, 2.6.1 and 2.6.2, Theorem 3.1 p. 27 in
[Bil99] and Lemma 2.A.3.

2.A Some technical results

Let us state first a Grönwall-type lemma which has been used to get moment estimates.
The proof can be found in [GL21a].
Lemma 2.A.1 (Grönwall-type lemma). Fix r ∈ [0, 1) and t0 ∈ R. Assume that g is a
non-negative real-valued function, b is a positive function and a is a differentiable real-
valued function. Moreover, suppose that the function bg r is a continuous function.

118
 2.A. Some technical results

Assume that Z t
∀t ≥ t0 , g(t) ≤ a(t) + b(s)g(s)r ds. (2.55)
t0
1
Then, setting Cr := 2 1−r ,
 1

 Z t 
1−r
∀t ≥ t0 , g(t) ≤ Cr a(t) + (1 − r) b(s) ds .
t0

Remark 2.A.2. Call H the right-hand side of (2.55). If g is not continuous, note that
the function H is still continuous and satisfies (2.55) (since b is positive and g ≤ H). So,
one can apply the lemma to H and thereafter use the inequality g ≤ H.

We state now a technical lemma about the convergence in the spaces C and D. We
recall that the space of continuous functions C is endowed with the uniform topology

+∞
1
 
du : (f, g) ∈ C((0, +∞), R)2 7→
X
n
min 1, sup |f − g| .
n=1 2 [ 1 ,n] n

Set

Λ := {λ : R+ → R+ , continuous and increasing function s.t. λ(0) = 0, lim λ(t) = +∞}

t→+∞

and 
1




1 if n
≤ t ≤ n,

kn (t) = n+1−t if n < t < n + 1,



0 if n + 1 ≤ t.


The space of càdlàg functions D is endowed with the Skorokhod topology ds defined for
(f, g) ∈ D((0, +∞))2 by
  
+∞
X 1   λ(t) − λ(s) 
n
1 ∧ inf a, ∃λ ∈ Λ, sup log ≤ a, sup |kn (t) (f ◦ λ(t) − g(t))| ≤ a .
n=1 2  s6 =t t − s t≥ 1 
n

Lemma 2.A.3.
(i) The uniform distance is finer than the Skorokhod one i.e. ds ≤ du .

119
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

(ii) Let (fε )ε≥0 , (hε )ε≥0 be two sequences of functions of D. If for all n ≥ 1,

P
sup |fε (t) − hε (t)| −→ 0,
1 ε→0
t∈[ n ,n]

P
then d(fε , hε ) −→ 0, where d ∈ {du , ds }.
ε→0
Proof. Let f, g be two càdlàg functions. The first point is true using the definition of the
metrics ds and du and noting that
 
 λ(t) − λ(s) 
inf a, ∃λ ∈ Λ, sup log ≤ a, sup |kn (t) (f (λ(t) − g(t))| ≤ a
 s6=t t−s t≥ 1

n

≤ sup |f − g| .
1
[ n+1 ,n+1]

P
Let us now prove the second part. Assume that for all n ≥ 1, sup[ 1 ,n] |fε − hε | −→ 0,
n ε→0
+∞
as ε → 0. Fix η > 0 and choose N > 0 such that 1
≤ η2 . Then,
P
2n
n=N +1

N
η X 1
ds (fε , hε ) ≤ du (fε , hε ) ≤ + sup |fε − hε | .
2 n=1 2n [ 1 ,n]
n

 +∞ −1
η
It follows, by setting η 0 :=
P 1
2
,
2n
that
n=1

 
N
P  sup |fε − hε | > η 0 
X
P (d (fε , hε ) > η) ≤ −→ 0.
1 ε→0
n=1 [n ,n]

For the sake of completeness, we state and improve the result of Problem 4.12 p. 64
in [KS98], on a general metric space.
Lemma 2.A.4. Let S be a Polish metric space endowed with a Borel σ-field S. Suppose
that (Pn )n≥1 is a sequence of probability measures on (S, S) which converges weakly to
a probability measure P . Suppose, in addition, that the sequence (fn )n≥1 of real-valued
continuous functions on S is uniformly bounded and converges to a continuous function
f , the convergence being uniform on compact subsets of S. Then, we have
Z Z
lim fn (ω) dPn (ω) = f (ω) dP (ω).
n→+∞ S S

120
 2.A. Some technical results

Proof. Notice that, since (Pn )n≥1 converges weakly thus, it is tight. So, for each ε > 0,
there exists a compact subset K of S such that for any n ≥ 1, Pn (K) ≥ 1 − ε.
Let us decompose Z Z
fn dPn − f dP = A + B + C + D,
S S

where
Z Z Z Z Z
A := fn dPn , B := (fn − f ) dPn , C := f dPn , and D := f dPn − f dP.
S\K K S\K S S

Let M be a bound for the sequence (fn ). Thus, by the choice of K,

|A| ≤ M Pn (S \ K) ≤ M ε.

The third integral can be treated analogously. Besides, since the sequence (fn ) converges
uniformly on K to f , there exists nε such that for all n ≥ nε , supK |fn − f | ≤ ε. Thereby,
we get
|B| ≤ εPn (K).

The last integral is smaller than ε for n large enough, since Pn converges weakly to P ,
and this concludes the proof.

Remark 2.A.5. Lemma 2.A.4 could be applied with S = C([0, +∞)) or D([0, +∞)).
However, the result for S = C([0, +∞)) is already contained in Problem 4.12 p. 64 in
[KS98].

Lemma 2.A.6. Let (Yty )t≥0 be the solution to a time-homogeneous SDE driven by an
α-stable process, 
dYt = dLt + b(Yt ) dt

(2.56)
Y0 = y


The measurable function b is supposed to be such that (2.56) has a pathwise unique strong
solution. Then (Yt )t≥0 is a Markov process.
Namely, for any d ≥ 1, 0 ≤ t1 ≤ · · · ≤ td , u ≥ 0 and any bounded measurable function
φ : Rd → R,   h i
y y
E φ(Yt1 +u , · · · , Ytd +u ) Fu = E φ(Ytz1 , · · · , Ytzd ) y . (2.57)
z=Yu

Proof. We give a proof for d = 2, the general case being similar. Call (Yts,y ) the solution
to dYt = dLt + b(Yt ) dt, satisfying Ys = y. Let φ : R2 → R be a bounded measurable

121
Chapter 2 – Behavior of a time-inhomogeneous kinetic Lévy-driven model

function. Pick u ≥ 0. Consider, for y ∈ R and u ≤ s ≤ t the function

 Z s Z t 
G(y, s, t, u) := (Ysu,y , Ytu,y ) = y + Ls − Lu + b(Yh ) dh, y + Lt − Lu + b(Yh ) dh .
u u

y y
Pick 0 ≤ s ≤ t. Using pathwise uniqueness, (Ys+u , Yt+u ) = G (Yuy , s + u, t + u, u). More-
u,y
over, by time-homogeneity of the SDE, (Ys+u )s≥0 and (Ysy )s≥0 have the same distribution.
As a consequence, G(y, s + u, t + u, u) = G(y, s, t, 0). Besides, by Markov property of Lévy
processes, the function G is independent of Fu . Hence,
     
y y
E φ(Ys+u , Yt+u ) Fu = E φ ◦ G (Yuy , s + u, t + u, u) Fu = E φ ◦ G (z, s, t, 0) Fu
z=Yuy

= E [φ ◦ G (z, s, t, 0)]z=Yuy = E [φ(Ysz , Ytz )]z=Yuy

This concludes the proof.

Let us now state and prove the following result.

Lemma 2.A.7. Let b be a function such that limt→+∞ b(t) = 0. Pick a > 0, γ ≥ 1 and
κ > 1. Let g be a continuously differentiable positive function satisfying

κ+γ−1 κ−1
g 0 (t) ≤ −ag(t) κ + b(t)g(t) κ , t ≥ 0. (2.58)

Then, g(t) −→ 0.
t→+∞
γ
Proof. Pick ε > 0. Let t1 be a positive real such that for all t ≥ t1 , |b(t)| ≤ a2 ε κ .
Step 1. We first show that there exists t∗ ≥ t1 , such that g(t∗ ) ≤ ε.
Assume, by way of contradiction, that it is not the case. Thus, one can consider the
1
function y = g κ , which satisfies

κy 0 (t) ≤ −ay(t)γ + b(t), t ≥ t1 . (2.59)

For all t ≥ t1 , we have

γ a γ a γ
κy 0 (t) ≤ −aε κ + ε κ ≤ − ε κ .
2 2
As a consequence, for all t ≥ t1 ,

1 a γ
κε κ < κy(t) ≤ κy(t1 ) − (t − t1 ) ε κ −→ −∞.
2 t→+∞

This is a contradiction.

122
 2.A. Some technical results

Step 2. We show that for all t ≥ t∗ , g(t) ≤ ε.

Define T = inf{t ≥ t∗ , g(t) > ε}. By continuity of the function g, we have g(T ) = ε.
Hence,
κ+γ−1 a γ κ−1 a κ+γ−1
g 0 (T ) ≤ −aε κ + ε κ ε κ < − ε κ < 0.
2 2
Therefore, there exists δ0 > 0, such that for all 0 < δ ≤ δ0 ,

g(T + δ) < g(T ) = ε.

This is a contradiction with the definition of T .

Acknowledgements
The authors would like to thank Nicolas Fournier and Thomas Cavallazzi for helpful
discussions about moment estimates. We would also like to thank Vlad Bally for his
suggestion which gives rise to the Section 2.6. We are grateful to Thomas Cavallazzi for
his careful reading of the manuscript.

123
 Part II

Time-inhomogeneous stochastic
dynamics in the quadratic potential

125
CONTENTS

Contents

Introduction 127

3 Behavior of the non-autonomous system driven by a Brownian motion133

3.1 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
3.2 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.3 Moment estimates and non-explosion . . . . . . . . . . . . . . . . . . . . . 136
3.4 Asymptotic behavior of the solution . . . . . . . . . . . . . . . . . . . . . . 137
3.4.1 Convergence of the f.d.d. in the super-critical regime . . . . . . . . 138
3.4.2 Convergence of the f.d.d. in the critical and sub-critical regimes . . 140
3.5 Proof of Corollary 3.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

4 Behavior of the non-autonomous system driven by a Lévy process 145

4.1 Notations and main results . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.1.1 Settings on Lévy processes and functional spaces . . . . . . . . . . . 145
4.1.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.2 Existence up to explosion . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.3 Moment estimates and non-explosion . . . . . . . . . . . . . . . . . . . . . 149
4.4 Asymptotic behavior of the solution . . . . . . . . . . . . . . . . . . . . . . 153
4.4.1 Convergence of the f.d.d. in the super-critical regime . . . . . . . . 155
4.4.2 Convergence of the f.d.d. in the critical and sub-critical regime . . . 158
4.5 Proof of Corollary 4.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

A Study of the deterministic underlying ODE 165

B Some technical results 168

126
 CONTENTS

Abstract: We consider a particle evolving in the quadratic potential, and subject to a time-
inhomogeneous frictional force and to a random force. The couple of its velocity and position is
solution to a stochastic differential equation driven by an α-stable Lévy process, and the frictional
force is of the form t−β sgn(v)|v|γ . We identify three regimes, depending on the balance between
β, γ and α, for the behavior in long-time of the couple velocity-position, with a suitable rescaling.

Keywords: kinetic stochastic equation; time-inhomogeneous stochastic differential equation;

Lévy process; scaling transformation; asymptotic distribution; asymptotic expansion of solu-
tions to ordinary differential equations.

MSC2020 Subject Classification: Primary 60H10; Secondary 60F17; 60G52; 60G18; 60J65;
34E05.

This part is a work in collaboration with Thomas Cavallazzi.

127
 I NTRODUCTION

In this part, we consider a particle, with velocity V ∈ R and position X ∈ R, evolving

2
in the quadratic potential U : x 7→ x2 , subject to a time-inhomogeneous frictional force
b and to a random force. This random force models the particle’s interaction with its
environment. Using Newton’s second law, the dynamics of the particle can be described
by the following stochastic damping Hamiltonian system, driven by an α-stable process
L with α ∈ (0, 2],





dVt = dLt − b(t, Vt ) dt − ∇U(Xt ) dt,

dXt = Vt dt, (SKEc )



(V , X ) = (v
0 , x0 ).

t0 t0

Stochastic Hamiltonian systems have been widely studied in the time-homogeneous

setting. An interesting problem is to understand their asymptotic behavior (see [AK94]).
The case of damping systems appears in [Wu01] (see also references therein).
The long-time behavior of a particle evolving in a free potential, i.e. when U = 0, has
already been studied, see e.g. [GO13], [FT21], [GL21a] and [GL21b]. When the random
force is supposed to be Brownian, i.e. α = 2, a non-linear Langevin’s type SDE with
time-homogeneous coefficients was studied in [FT21]. The system is of the form

ρZ t Z t
Vt = v0 + Bt − b(Vs ) ds and Xt = x0 + Vs ds,
2 0 0

v
with, for example, b(v) = 1+v 2 . Using a suitable rescaling, the authors show that the

position process behaves asymptotically as a Brownian motion, a stable process, or an

integrated Bessel process, depending on the value of ρ. For example, in the normal diffusive
regime, the rescaled velocity-position process behaves asymptotically as an ergodic process
and a Brownian motion. In the integrated Bessel regime, the Bessel process and its time
integral are the asymptotic distributions. However, the tools used therein, such as invariant
measure, scale function and speed measure, are limited to time-homogeneous coefficients.
In [GO13] and [GL21a], the drift coefficient b was allowed to depend on time under a
homogeneity condition. More precisely, in [GO13], the authors study the convergence in

128
distribution, when t tends to +∞, of rt Vt , for a certain rate of convergence rt . In [GL21a],
the authors extend to convergence of processes the results obtained in [GO13] for the cou-
ple velocity-position. Namely, the authors study the limit in distribution of the rescaled
process (rε,V Vt/ε , rε,X Xt/ε )t for two appropriate rates of convergence rε,V and rε,X . Results
were further generalized in [GL21b] to a Lévy driving process, supposed to be an α-stable
process. To be more precise, if b is of the form t−β sgn(v) |v|γ , the authors highlight three
schemes, depending on the balance between β, γ and α, the index of stability of L. When
the frictional force is sufficiently “small at infinity”, the rescaled process behaves as if
there was no frictional force and thus converges in distribution towards the Kolmogorov
process (L, 0· L), where L has the same distribution as the driving process. When the
R

two forces offset, the limiting process is still of kinetic form (V, 0· V), but the process
R

V is henceforth ergodic. Whereas, when the drag force swings with the random noise,
the limiting process is no longer kinetic. The rescaled velocity process converges in finite
dimensional distributions towards a product measure. The proofs are essentially based on
the self-similarity of the driving process and on moment estimates of the velocity process.

Degenerate systems has been intensively studied for several years. In particular, the
existence and uniqueness of solutions to degenerate SDEs have been discussed in many
papers. These models are called degenerate because the noise is only present in one com-
ponent of the system but can be transferred into others by drift terms. The well-posedness
of these systems, when their deterministic version is ill-posed, can be proved by taking
advantage of the regularizing effect of the noise and of its propagation through the whole
system. The case of Brownian degenerate SDEs has been of course wildly explored, see e.g.
[Fed+17], [WZ15], [Zha16], [HMC18]. The time-dependence is treated in the last three
cited papers. The case of a Lévy driving process is more recent, see e.g. [Zha14] for a
time-homogeneous setting, and [MM21] for drifts depending on time.

In this part, we are interested in the long-time behavior of the solution to (SKEc ),
driven either by a Brownian motion, i.e. α = 2, or by an α-stable Lévy process, with α ∈
(0, 2). We consider a time-inhomogeneous frictional force given by t−β sgn(v) |v|γ , where
β and γ are non-negative parameters. More precisely, our goal is to study the asymptotic
 
(ε)
behavior, as ε → 0, of the rescaled velocity-position process (Zt )t := rε (Xt/ε , Vt/ε )T ,
t
for an appropriate rate of convergence rε . One of our motivations is to study how the
presence of the quadratic potential influences the results obtained in [GL21a; GL21b]

129
Chapter 2 – Introduction

through a confining effect on the position X. Our system without noise and frictional
force is nothing else than the classical harmonic oscillator

v 0

= −xt ,
t
 x0

= vt .
t

The intrinsic oscillatory behavior induced by the quadratic potential prevents the rescaled
process Z (ε) from converging as a process. However, we will prove that each of its one-
dimensional marginal distributions, for example its distribution at time t = 1, converges
towards either a Gaussian distribution with independent coordinates in the Brownian
case or towards a symmetric α-stable distribution when α ∈ (0, 2). In order to obtain the
convergence of the whole process, the key idea is to remove the oscillations present in the
system. Namely, we set Yt := Θ−t (Xt , Vt )T , where Θ−t is the rotation on R2 of angle −t
defined by  
cos(t) sin(t)
Θ−t :=  .
− sin(t) cos(t)
β
As in [GL21a; GL21b], we highlight three regimes depending on the position of q := γ+α−1
1
with respect to α1 . The rate of convergence rε is given by rε := εq∧ α , where ∧ denotes
the minimum. Actually, the shift of regimes depends on the position of β with respect to
γ+α−1
α
. Consequently, the condition q > α1 should be read as β > γ+α−1 α
, but we keep the
notation q for the sake of simplicity, since we deal with parameters satisfying γ +α−1 > 0.
In the super-critical regime, i.e. q > α1 , we prove in Theorems 3.1.1 and 4.1.1 that the
(ε)
rescaled process (Yt )t := (rε Yt/ε )t converges in distribution towards either a standard
Brownian motion on R2 when α = 2, or a rotationally invariant stable process when α < 2.
In the critical regime, i.e. q = α1 , and in the sub-critical regime, i.e. q < α1 , we treat only
 i
the linear case γ = 1, and we assume that β ∈ 12 , 1 . In the critical regime, we prove
that the rescaled process converges in distribution, while in the sub-critical regime, we
obtain only the convergence in finite dimensional distributions. In every case, the limit is
explicit.
1
We deduce in Corollaries 3.1.2 and 4.1.2 the convergence in distribution of t−q∧ α Zt , as
t → +∞. Let us mention that the rate of convergence of the position X is different from
that found in [GL21a; GL21b], when U = 0. Indeed, contrary to the free potential system,
the position process is somehow more diffusive. This is due to the structure of our model.
Namely, the presence of the quadratic potential allows the noise to propagate more effi-

130
ciently from the velocity component to the position one (see [Fed+17] for more details).
This explains why both the limiting processes and the rate of convergence are different
between our work and [GL21a; GL21b]. Let us also note that the position process grows
more slowly in our case than when U = 0. For example, in the Brownian super-critical
3
regime, Xt behaves asymptotically as N (0, 2t ) in our framework, but as N (0, t3 ) in the
free potential one. This difference can also be seen in moment estimates established for
the position process X (see Remarks 3.3.2 and 4.3.2). This can be explained by the fact
that the quadratic potential confines the particle through a spring force. Let us also no-
tice that in the Brownian case, the velocity and position processes become asymptotically
independent contrary to the α-stable case (see Remark 4.1.3).

In our model, the particle is no longer free. Consequently, both equations are intrinsi-
cally linked to each other. Therefore, we can no longer separate by components the study
of the velocity-position process. We thus write the system (SKEc ) in a vector viewpoint,
as done in [Fed+17], and use a variation of constants method to bring back to the study of
a two-dimensional system in a free potential. We then adapt the methods used in [GL21a;
GL21b]. The behavior depends again on the balance between the drag force and the ran-
dom force. In the super-critical regime, the proof essentially relies of the self-similarity of
the driving process and on moment estimates of V and X. In the critical and sub-critical
regimes, we need to restrict ourselves to a linear drag force, i.e. γ = 1, in order to rely on
the study of the asymptotic behavior of the underlying non-autonomous ordinary differ-
ential equation (ODE).
When the driving process is Brownian, we take advantage of the theory of Gaussian pro-
cesses. The convergence is thus characterized by the study of the mean and covariance
functions. In the case of a stable driving process, we need to study the convergence, in
distribution and as a process, of a Wiener-Lévy integral (i.e. the integral of a deterministic
function integrated against a stable process). The key point here is to use the fact that a
Wiener-Lévy integral is a process with independent increments.

We consider the case of a Brownian driving process in Chapter 3 and we follow the
same structure for a pure-jump α-stable driving process in Chapter 4. For the sake of
clarity, we opt for separating the two cases since the tools used are different. In each
chapter, we introduce some notations and state our main results. We start studying the
existence of the solution to (SKEc ) and then give estimates of the moment. The proofs of

131
Chapter 2 – Introduction

the main results are then presented. Finally, we state and prove some technical results in
appendix.

Throughout this part, for x, y ∈ R2 , kxk represent the Euclidean norm of x, and x · y
the inner product of x and y. If x ∈ R2 , for each i ∈ {1, 2}, x(i) denotes its i-th component.
We call I2 the identity matrix of dimension 2 and AT is the transpose matrix of a matrix
A. We denote by C some positive constant, which may change from line to line, and we
use subscripts to indicate the parameters on which it depends when it is necessary.

132
 Chapter 3

B EHAVIOR OF THE NON - AUTONOMOUS

DEGENERATE SYSTEM IN THE
QUADRATIC POTENTIAL , DRIVEN BY A
B ROWNIAN MOTION

3.1 Notations and main results

Throughout the chapter, to keep standard notations, we denote by B a standard
Brownian motion The space of continuous functions C((0, +∞), R) is endowed with the
uniform metric on compact subsets

+∞
1
 
2
X
du : (f, g) ∈ C((0, +∞), R) →
7 n
1 ∧ sup |f − g| ,
n=1 2 [ 1 ,n] n

where ∧ denotes the minimum. For simplicity, we shall write C for C((0, +∞), R).

Let us fix t0 > 0, β ≥ 0, γ > 0 and (v0 , x0 ) ∈ R2 . We consider the following system of
SDEs, defined on the time interval [t0 , +∞),

|Vt |γ

dVt = dBt − sgn(Vt ) β dt − Xt dt,



t



 dX t = Vt dt, (3.1)



(V
t0 , Xt0 ) = (v0 , x0 ).


The previous system can be written in a vector viewpoint. Indeed, we set, for all t ≥ t0

133
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

and v ∈ R,
         
Xt 0 0 1 0 0 0
Zt :=   , Wt :=   , A :=  , Γ :=  and F (t, v) :=  |v|γ  .
 
Vt Bt −1 0 0 1 sgn(v) β
t

Thereby, the system (3.1) can be rewritten as


dZt

= Γ dWt + AZt dt − F (t, Vt ) dt,
(3.2)
Zt = z0 := (x0 , v0 ).

0

π
Notice that the matrix A is the rotation matrix of angle 2
and that, for all t ∈ R,
 
cos(t) sin(t) 
Θ−t := etA = ,
− sin(t) cos(t)

which is the rotation matrix of angle −t.

To state the main results of our paper, we define, for any t ≥ t0 , Yt := e−tA Zt . We
easily check, with Itô’s formula, that Y is given by

dYt = e−tA Γ dSt − e−tA F (t, Vt ) dt. (SDEY )

We are now able to state our results on the long-time behavior of these systems.
β 1 (ε)
Theorem 3.1.1. Define q := γ+1 , rε := εq∧ 2 and set (Yt )t≥εt0 := (rε Yt/ε )t≥εt0 . Let B be
a standard Brownian motion on R2 .
(i) (Super-critical regime i.e. 2q > 1). The rescaled process Y (ε) converges in distribution
 
in C towards B 2t .
t>0
(ii) (Critical regime i.e. 2q = 1). Assume that γ = 1. The rescaled process Y (ε) converges
 R √ 
in distribution in C towards the Gaussian process √12t 0t s dBs .
t>0
1
(iii) (Sub-critical regime i.e. 2q < 1). Assume that γ = 1 and β > The rescaled process
2
.
Y (ε) converges in finite dimensional distributions towards the centered Gaussian
process with covariance kernel K(s, t) = 12 sβ 1{s=t} I2 .

The question of the convergence of the rescaled process associated with Z is treated
in the following corollary.

134
 3.2. Existence up to explosion

(ε) 1
Corollary 3.1.2. Let us define (Zt )t≥εt0 := (rε Zt/ε )t≥εt0 , where rε := εq∧ 2 . The rescaled
process Z (ε) does not converge in distribution. However, under the assumptions of the
preceding theorem, we deduce from Theorem 3.1.1 the convergence in distribution of r1/t Zt
towards explicit limits, as t → +∞.
The limit is either N (0, 21 I2 ) in the super-critical and sub-critical regimes, or N (0, 41 I2 )
in the critical regime.

Remark 3.1.3. The rate of convergence of the position process is the same as for the ve-
locity process. This is not true in the free potential case, i.e. when U = 0 (see [GL21a]). In-
deed, our rate of convergence for X is smaller. For example in the Brownian super-critical
regime, the variance of Xt heuristically behaves as 2t in large time, while in [GL21a], it
3
behaves as t3 . This is quite natural because the quadratic potential tends to confine the
particle through a spring force, so the particle spreads out more slowly than without po-
tential.

3.2 Existence up to explosion

Theorem 3.2.1. The system of SDEs (3.1) admits a unique (global) strong solution if
γ ∈ (0, 1]. And if γ > 1, there exists a unique strong solution defined up to its explosion
time τ∞ .

Proof. In the case γ > 1, the coefficients of the SDE (3.2) are locally Lipschitz continuous
with respect to the space variable, locally uniformly in time. Thus, by Theorem 21.3 p.
415 in [Kal02], there exists a unique solution up to explosion. The argument is standard,
see [SG03] for details.
Assume now that γ ≤ 1. We will use Theorem 1 in [HMC18]. Keeping the same notations,
we have in our case, for any (x, v) ∈ R2 and t ≥ t0 , F1 (t, v, x) := − sgn(v)|v|γ t−β − x,
F2 (t, v, x) := v and σ(t, v, x) = 1. Assumptions (ML) and (UE) in [HMC18] are obviously
satisfied. Let us now remark that F1 is γ-Hölder with respect to v ∈ R uniformly with
respect to t ≥ t0 and x ∈ R, and is Lipschitz continuous with respect to x, uniformly with
respect to t and v. With the notations used in [HMC18], we have β1 = γ and β2 = 1.
Thus, Assumption (Tβ ) is satisfied. Finally, we check that Assumption (Hη ) is satisfied.
Since ∂v F2 = 1, we can conclude, taking η small enough and E1 = {1}, that there exists
a unique strong solution to (3.1).

135
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

3.3 Moment estimates and non-explosion

In this section, we state and prove moment estimates of Z. It will be useful to control
some stochastic terms appearing later. For all n ≥ 0, define the stopping time

τn := inf{t ≥ t0 , kZt k ≥ n}.

Set τ∞ := limn→+∞ τn the explosion time of Z.

Proposition 3.3.1. The explosion time of Z is a.s. infinite and, for all κ ≥ 0,

κ
E [kZt kκ ] ≤ Cκ,t0 t 2 . (3.3)

Remark 3.3.2. Let us mention that the moment estimate obtained for the position process
X is a priori smaller in our case than in the free potential case [GL21a]. It is explained
by the confining effect of the quadratic potential.

Proof. The proof is adapted from [GL21a] to two-dimensional processes. For the sake of
completeness, we sketch the proof in our context.
Using Itô’s formula applied to the function f : (x, v) 7→ x2 + v 2 and the fact that for all
z ∈ R2 , z · Az = 0, we deduce that, for all t ≥ t0 ,
Z t Z t∧τn
kZt∧τn k2 ≤ kz0 k2 + 21{s≤τn } Zs · (Γ dWs ) − 2Zs · F (s, Vs ) ds + (t − t0 ).
t0 t0

Remark that for any s ≥ t0 , Zs · F (s, Vs ) = Vs sgn(Vs ) |Vs |γ s−β ≥ 0. Taking expectation
yields
h i
E kZt∧τn k2 ≤ kz0 k2 + (t − t0 ) ≤ Ct0 t.

Thanks to Lemma B.0.1, we can conclude that the explosion time of Z is a.s. infinite. Set
κ ∈ [0, 2], so, by Jensen’s inequality and Fatou’s lemma

h iκ  h i κ2 κ
E [kZt∧τ∞ kκ ] ≤ E kZt∧τ∞ k2 2
≤ lim inf E kZt∧τn k2
n→∞
≤ Cκ,t0 t 2 . (3.4)

This leads to (3.3).

136
 3.4. Asymptotic behavior of the solution

When κ > 2, v 7→ kvkκ is a C 2 -function, so by Itô’s formula, for all t ≥ t0 ,

Z t∧τn Z t∧τn
kZt∧τn kκ ≤ kz0 kκ + κ kZs kκ−2 Zs · (Γ dWs ) − κ kZs kκ−2 Zs · F (s, Vs ) ds
t0 t0
Z t∧τn
+ Cκ kZs kκ−2 ds.
t0

In addition, it follows from the hypothesis on the sign of the drift function that
Z t Z t∧τn
kZt∧τn kκ ≤ kz0 kκ + κ1{s≤τn } kZs kκ−2 Zs · (Γ dWs ) + Cκ kZs kκ−2 ds. (3.5)
t0 t0

Taking expectation in (3.5), we have

Z t h i
κ κ κ
E [kZt∧τ∞ k ] ≤ lim inf E [kZt∧τn k ] ≤ kv0 k + Cκ E kZs kκ−2 ds.
n→∞ t0

h i
When 0 ≤ κ − 2 ≤ 2, we can upper bound E kZs kκ−2 by injecting (3.4) and get
Z t
κ−2 κ
E [kZt∧τ∞ kκ ] ≤ kv0 kκ + Cκ,t0 s 2 ds ≤ Cκ,t0 s 2 .
t0

The method is then applied inductively to prove the inequality for all κ > 2.

3.4 Asymptotic behavior of the solution

We gather in this section the proof of Theorem 3.1.1. The strategy is to prove the
convergence of the finite dimensional distributions (f.d.d.) of the process Y (ε) , as ε → 0,
and its tightness in the critical and super-critical regimes. We first focus on the tightness.
n √ o
Lemma 3.4.1. If 2q ≥ 1, then the family ( εYt/ε )t≥εt0 , ε > 0 is tight on every compact
interval [m, M ], with 0 < m ≤ M .

Proof. We use the Kolmogorov criterion stated in Problem 4.11 p. 64 in [KS98].

Take ε0 small enough such that for all ε ≤ ε0 , we have εt0 ≤ m. Fix m ≤ s ≤ t ≤ M and
a > 4. Define, for t ≥ εt0 , the local martingale term appearing in (SDEY )
 
(ε) √ Z t/ε
−sA
√ Z t/ε − sin(s)
Mt := ε e Γ dWs = ε  dBs . (3.6)
t0 t0 cos(s)

137
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

Using Jensen’s inequality, moment estimates (see Proposition 3.3.1) and Burkholder-
Davis-Gundy’s inequality (see Theorem 4.4.22 p. 263 in [App09]), we have
a#
√ Z
"
h ai h ai t/ε
(ε) (ε) −uA
E Yt − Ys(ε) ≤ Ca E Mt − Ms(ε) + Ca E ε e F (u, Vu ) du
s/ε
"Z #
ai t/ε
1− a2
h
(ε) a−1 a
≤ Ca E Mt − Ms(ε) + Ca ε (t − s) E kF (u, Vu )k du
s/ε
 E a/2  Z t/ε
D a γa
≤ Ca E Tr M·(ε) − Ms(ε) + Ca ε1− 2 (t − s)a−1 u 2 −βa du
t s/ε
a
a(β− γ+1
≤ Ca (t − s) + Ca,m,M ε
2 2
)
(t − s)a−1
a
≤ Ca,m,M (t − s) 2 .

Since a2 > 2 and β ≥ γ+1

2
the upper bound is independent of ε ≤ 1. Furthermore, by
moment estimates (Proposition 3.3.1),
h i √
sup E Ym(ε) ≤ m < ∞.
ε≤ε0

Thus, Kolmogorov’s criterion can be applied, proving the tightness result.

We will now prove the convergence of the finite-dimensional distributions of Y (ε) .

Thanks to the previous lemma, this will yield the weak convergence on every compact
set (see Theorem 13.1 p. 139 in [Bil99]). The convergence in distribution on the whole
space C will follow, for 2q ≥ 1, from Theorem 16.7 p. 174 in [Bil99], since all processes
considered are continuous.

3.4.1 Convergence of the f.d.d. in the super-critical regime

(ε) √
Assume here that 2q > 1. Recall that (Yt )t≥εt0 := ( εYt/ε )t≥εt0 .

Proof of Theorem 3.1.1 (i).

Step 1. We first prove the convergence of the f.d.d. of the local martingale term M (ε)
appearing in (SDEY ).
Recall that the stochastic integral M (ε) was defined in (3.6). It is a centered Gaussian
process with covariance kernel defined, for any (s, t) ∈ [εt0 , +∞)2 , by
 
(ε)
Cov(Ms(ε) ) Cov(Ms(ε) , Mt )
K (ε) (s, t) :=  (ε) (ε) ,
Cov(Mt , Ms(ε) ) Cov(Mt )

138
 3.4. Asymptotic behavior of the solution

where  
(ε,1) (ε,2)
(ε) Cov(Ms(ε,1) , Mt ) Cov(Ms(ε,1) , Mt )
Cov(Ms(ε) , Mt ) =  (ε,1) (ε,2) ,
Cov(Ms(ε,2) , Mt ) Cov(Ms(ε,2) , Mt )

and Cov(Ms(ε) ) = Cov(Ms(ε) , Ms(ε) ). Thus, the convergence of the f.d.d. of M (ε) reduces on
the study of the limit of K (ε) , when ε converges to 0. Let us fix εt0 ≤ s ≤ t. Using that
M (ε) has independent increments and by Itô’s isometry, we find that
 R s/ε R s/ε 
2
(ε) ε t sin(u) du −ε t0 sin(u) cos(u) du
Cov(Ms(ε) , Mt ) =  R s/ε0 R s/ε .
−ε t0 sin(u) cos(u) du ε t0 cos(u)2 du

We get that, for all 0 < s ≤ t,

 
(ε) 1 s 0
Cov(Ms(ε) , Mt ) −→ .
ε→0 2 0 s
 
We recognize the covariance kernel of the process B 2t , where B denotes a standard
t>0
2
Brownian motion on R . Since mean and covariance functions characterize Gaussian pro-
(ε)
cess (see Lemma 13.1 (i) p. 250 in [Kal02]), we have thus proved that (Mt )t≥εt0 converges
 
in f.d.d. towards B 2t .
t>0
Step 2. Pick T > 0. We prove that
" #
(ε) (ε)
E sup Yt − Mt −→ 0.
εt0 ≤t≤T ε→0

We have
(ε) (ε) √ √ Z T /ε
sup Yt − Mt ≤ ε kz0 k + ε e−sA F (s, Vs ) ds.
εt0 ≤t≤T t0

We use moment estimates (Proposition 3.3.1) to get

√ Z √ Z
" # " #
T /ε T /ε
−sA
E ε e F (s, Vs ) ds = E ε kF (s, Vs )k ds
t0 t0

√ Z
" #
T /ε
γ −β
≤E ε |Vs | s ds
t0
√ Z T /ε
γ
≤ εCκ,t0 s 2 −β ds
t0
β− γ+1 γ √ γ2 −β+1
≤ Cκ,t0 (ε 2 T 2 −β+1 − εt0 ).

139
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

γ+1 1
Hence, setting r := min(β − 2
, 2 ), which is positive by assumption, we get
" #
(ε) (ε)
E sup Yt − Mt = O (εr ).
εt0 ≤t≤T ε→0

We conclude the proof by using Theorem 3.1 p. 27 in [Bil99].

3.4.2 Convergence of the f.d.d. in the critical and sub-critical

regimes
 i
1
In this section, we consider the linear case, i.e. γ = 1. Pick β ∈ 2
,1 . Recall that

(ε)
(Yt )t≥εt0 := (εq Yt/ε )t≥εt0 ,

β
where q = γ+1
.

Proof of Theorem 3.1.1 (ii) and (iii). Leaving out the Brownian term, the underlying
ODE of our system is the following

x0 (t)
x00 (t) + + x(t) = 0, t ≥ t0 . (3.7)
tβ

Pick the basis of solutions given in Lemma A.0.3 and denote by R its resolvent matrix.
By Itô’s formula, we get, for all t ≥ t0 ,
Z t
Yet := Rt−1 Zt = Rt−1
0
Z0 + Rs−1 Γ dWs .
t0

Let us define f the rate of decrease of R (see Lemma A.0.3) by


 √1

if β = 1,
∀t > 0, f (t) :=  t  1−β

exp − t else.
2(1−β)

Set, for t ≥ εt0 ,

e−tA Rt
  Z t/ε
f (ε) t
Φt := and M t
q
:= ε f Rs−1 Γ dWs . (3.8)
f (t) ε t0

140
 3.4. Asymptotic behavior of the solution

Pick t ≥ εt0 . To study the convergence of Y (ε) we decompose it into

t
 
(ε) f (ε) .
Yt =ε fq
Φt/ε Rt−1 Z0 + Φt/ε Mt (3.9)
ε 0

 
t
Let us notice that, for any t > 0, εq f ε
converges to 0, as ε → 0.

Step 1. We first note the convergence of Φ.

Using the asymptotic expansion of the resolvent matrix (Lemma A.0.3), we can write, for
t ≥ εt0 ,
 
Φt = I2 + O t1−2β .
t→∞

As a consequence, since 1 − 2β < 0, Φt/ε converges to the identity matrix I2 , as ε → 0.

Moreover, we obtain that

t
 
q
εf Φt/ε Rt−1 Z0 −→ 0.
ε 0 ε→0

As a consequence, we can forget the first term appearing in the decomposition (3.9) of
Y (ε) (see Theorem 3.1 p. 27 in [Bil99]).
Step 2. We compute now the covariance kernel of Y (ε) defined in (3.9).
It is defined, for (s, t) ∈ [εt0 , +∞)2 , by
 
(ε)
(ε) Cov(Ys(ε) ) Cov(Ys(ε) , Yt )
K (s, t) :=  (ε) (ε) ,
Cov(Yt , Ys(ε) ) Cov(Yt )

where  
(ε,1) (ε,2)
(ε) Cov(Ys(ε,1) , Yt ) Cov(Ys(ε,1) , Yt )
Cov(Ys(ε) , Yt ) = (ε,1) (ε,2) ,
Cov(Ys(ε,2) , Yt ) Cov(Ys(ε,2) , Yt )

and Cov(Ys(ε) ) = Cov(Ys(ε) , Ys(ε) ).

Moreover, we get, for all (s, t) ∈ [εt0 , +∞)2 ,
 
f (ε) , Φ M (ε) f )ΦT .
f (ε) , M (ε)
Cov Φs/ε Ms
f
t/ε t = Φs/ε Cov(Ms t t/ε

Using the expression of the Wronskian obtained in Lemma A.0.3, we obtain, for all t ≥ εt0 ,
 
(ε)
Z t/ε
−y2 (u)
M
f
t = εq f (t/ε) f (u)−2  dBu .
t0 y1 (u)

141
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

It is a centered Gaussian process and for any εt0 ≤ s ≤ t, we have

 
(ε)
Z s/ε
y22 (u) −y2 (u)y1 (u)
Cov(Ms
f ) = εβ f (t/ε)f (s/ε)
f (ε) , M
t f (u)−4  du.
t0 −y2 (u)y1 (u) y12 (u)

Using the asymptotic expansion of the solutions and Lemma B.0.3, we can write, for all
εt0 < s ≤ t,
 
(ε)
Z s/ε
sin2 (u) − sin(u) cos(u)
Cov(Ms
f ) = εβ f (t/ε)f (s/ε)
f (ε) , M
t f (u)−2  du
t0 − sin(u) cos(u) cos2 (u)
 
+ O ε2β−1 f (t/ε)f (s/ε)−1 .
ε→0

Moreover, using asymptotic expansions of these integrals (see Lemmas B.0.2 and B.0.3),

Z s/ε
1 Z s/ε
εβ f (t/ε)f (s/ε) f (u)−2 cos2 (u) du = εβ f (t/ε)f (s/ε) f (u)−2 du
t0 2 t0
 
+ o f (t/ε)f (s/ε)−1 .
ε→0

The same equality holds for

Z s/ε
β
ε f (t/ε)f (s/ε) f (u)−2 sin2 (u) du,
t0

and we have
Z s/ε  
β
ε f (t/ε)f (s/ε) f (u)−2 cos(u) sin(u) du = o f (t/ε)f (s/ε)−1 .
t0 ε→0

Thanks to Lemma B.0.3, this leads to

" #
f (ε) ) 1 Z s/ε  
f (ε) , M
Cov(M s t = εβ f (t/ε)f (s/ε) f (u)−2 du I2 + o f (t/ε)f (s/ε)−1
2 t0 ε→0

f (t/ε) β  
= kβ s I2 + o f (t/ε)f (s/ε)−1 ,
f (s/ε) ε→0

where 
1

if β = 1,
4
kβ :=
1

else.
2

Thus, we have proved the convergence of the f.d.d. of Y (ε) . Note that when 2q = 1,

142
 3.5. Proof of Corollary 3.1.2

 Rt √ 
we recognize the covariance kernel of the process √1 s dBs , where B denotes a
2t 0 t>0
standard Brownian motion on R2 .

Remark 3.4.2. The proof relies on the asymptotic expansion of the resolvent matrix of
 i
(3.7). We were able to prove it only for β ∈ 12 , 1 . However, if β = 0, the resolvent
 
matrix is explicit and following the same lines, we can prove that Zt/ε converges
t≥εt0
in f.d.d. towards a centered Gaussian process with covariance kernel (s, t) 7→ 12 I2 1{s=t} .
This behavior can be explained by the fact that the frictional force does not decrease along
time. This cancels somehow the rotation bearing, which prevents Z (ε) from converging as
a process when β > 0.

3.5 Proof of Corollary 3.1.2

Proof of Corollary 3.1.2. We start by proving the convergence in distribution of v1/T ZT ,
as T → +∞. We claim that it follows from Theorem 3.1.1. Indeed, it is enough to
remark that the convergence results stated in Theorem 3.1.1 imply the convergence in
distribution of the marginal distribution at time t = 1 of Y (ε) . Let us also recall that
ZT = eT A YT . Setting T = 1ε , the convergence of v1/T ZT is therefore a direct consequence
of Lemma B.0.4.
We now show that the rescaled process Z (ε) does not converge in distribution. We do the
proof only in the super-critical regime. Assume by contradiction that it is the case. Hence,
each of its coordinates shall converge too. We thus have the convergence of the rescaled
process X (ε) . Using (SDEY ), we can write

√ √ √ Z t/ε 
t

√ Z t/ε t
 
εXt/ε = εx0 + ε sin − s dBs − ε sin − s F (Vs )s−β ds.
t0 ε t0 ε

As in the proof of Theorem 3.1.1, the last term converges in probability uniformly on
compact intervals towards zero. Hence, the following term shall converge in distribution

(ε) √ Z t/ε 
t

It := ε sin − s dBs .
t0 ε
(ε)
The process (It )t≥εt0 is Gaussian, thereby its limit shall be Gaussian too and its covari-
ance function shall converge (see Lemma 13.1 (i) p. 250 in [Kal02]). However, using Itô’s

143
Chapter 3 – Behavior of the non-autonomous system driven by a Brownian motion

isometry, one can compute, for εt0 ≤ s ≤ t,

Z s/ε
t s
h i    
(ε)
E It Is(ε) =ε sin − u sin − u du
t0 ε ε
1 t−s s 1 t−s t−s
        
= ε cos − t0 + sin − sin − 2t0
2 ε  ε 2 ε ε
1 t−s

= s cos + o (1).
2 ε ε→0

This term does not converge if s 6= t, and that concludes the proof.

144
 Chapter 4

B EHAVIOR OF THE NON - AUTONOMOUS

DEGENERATE SYSTEM IN THE
QUADRATIC POTENTIAL AND DRIVEN BY
A L ÉVY PROCESS

4.1 Notations and main results

4.1.1 Settings on Lévy processes and functional spaces
Throughout the paper, we denote by L a symmetric α-stable Lévy process on R with
α ∈ (0, 2]. We call ν its Lévy measure, which can be written as ν(dz) = a |z|−1−α 1{z6=0} dz
with a > 0. As a Lévy measure, it satisfies R∗ (1 ∧ z 2 )ν(dz) < +∞. Let us also denote
R

by ψ the characteristic exponent of the stable process L. It follows from Theorem 14.15
p. 86 in [Sat99] that, for all ξ ∈ R,

ψ(ξ) = −a|ξ|α . (4.1)

We denote by N the Poisson random measure associated with L and by N f its compensated

Poisson measure. Using Lévy-Itô’s decomposition, we have, for all t ≥ 0,

Z t Z


 zN (ds, dz) if α ∈ (0, 1),
∗
Z0 t ZR


 Z tZ
Lt = z N (ds, dz) +
f zN (ds, dz) if α = 1,

 0 Z{0<|z|<1} 0 {|z|≥1}

 Z t
if α ∈ (1, 2).
 f(ds, dz)

 zN
0 R∗

145
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

Set Λ := {λ : R+ → R+ , continuous and increasing s.t. λ(0) = 0, lim λ(t) = +∞} and
t→+∞


1




1 if n
≤ t ≤ n,

kn (t) = n+1−t if n < t < n + 1,



0 if n + 1 ≤ t.


The space of right-continuous with left limits (càdlàg) functions D((0, +∞), R), i.e. the
functions defined on (0, +∞) which are càdlàg on every compact subinterval of (0, +∞),
is endowed with the Skorokhod metric ds defined for (f, g) ∈ D((0, +∞), R)2 by
  
+∞
X 1   λ(t) − λ(s) 
n
1 ∧ inf a, ∃λ ∈ Λ, sup log ≤ a, sup |kn (t) (f ◦ λ(t) − g(t))| ≤ a .
n=1 2  s6=t t − s t≥ 1 
n

For simplicity, we shall write D for D((0, +∞), R).

4.1.2 Main results

Let us fix t0 > 0, β ≥ 0, γ > 0 and (v0 , x0 ) ∈ R2 . We consider the following system of
SDEs, defined on the time interval [t0 , +∞),

|Vt |γ

dVt = dLt − sgn(Vt ) β dt − Xt dt,


t



dXt = Vt dt, (SKE)




(V
t0 , Xt0 ) = (v0 , x0 ).


Following the same lines as in the previous chapter, we rewrite (SKE) as


dZt

= Γ dSt + AZt dt − F (t, Vt ) dt,
(4.2)
Zt = z0 := (x0 , v0 )T ,

0

where for all t ≥ t0 and v ∈ R,

         
Xt 0 0 1 0 0 0
Zt :=   , St :=   , A :=  , Γ :=  and F (t, v) :=  |v|γ  .
 
Vt Lt −1 0 0 1 sgn(v) β
t

To state the main results of our paper, we define, for any t ≥ t0 , Yt := e−tA Zt , which

146
 4.1. Notations and main results

is given by
dYt = e−tA Γ dSt − e−tA F (t, Vt ) dt. (SDEY )

We are now able to state our results on the long-time behavior of these systems.
β 1
Theorem 4.1.1. Assume that γ ∈ (0, α). Define q := γ+α−1 , rε := εq∧ α and set
(ε)
(Yt )t≥εt0 := (rε Yt/ε )t≥εt0 . Let L be a rotationally invariant stable process on R2 , whose
characteristic exponent is given by

a Z 2π
ξ ∈ R 7→ 2
−Ce kξkα , with Ce := |cos(x)|α dx.
2π 0

(i) (Super-critical regime i.e. αq > 1). Assume that α ∈ (1, 2). The rescaled process
Y (ε) converges in distribution in D towards (Lt )t>0 .
(ii) (Critical regime i.e. αq = 1). Assume that γ = 1. The rescaled process Y (ε) converges
 R √ 
in distribution in D towards the Lévy-type process √1t 0t s dLs .
t>0
(iii) (Sub-critical regime i.e. αq < 1). Assume that γ = 1 and β > 12 . Then, for all
 
(ε) (ε)
(t1 , · · · , td ) ∈ (0, +∞)d , Yt1 , · · · , Ytd converges in distribution towards the prod-
uct measure µt1 ⊗ · · · ⊗ µtd , where µt is the distribution with characteristic function

2
 
7 exp − Ce kξkα tβ .
ξ∈R → 2
α

The question of the convergence of the rescaled process associated with Z is treated
in the following corollary.
(ε) 1
Corollary 4.1.2. Let us define (Zt )t≥εt0 := (rε Zt/ε )t≥εt0 , where rε := εq∧ α . The rescaled
process Z (ε) does not converge in distribution. However, under the assumptions of the
preceding theorem, we deduce from Theorem 4.1.1 the convergence in distribution of r1/t Zt
towards explicit limits, as t → +∞.
Keeping the same notations as in the preceding theorem, the characteristic function of the
limit is given, for all ξ ∈ R2 , by
 
(i) exp −Ce kξkα in the super-critical regime,
  −1 
(ii) exp − 1 + α
2
Ce kξkα in the critical regime,
 
(iii) exp − α2 Ce kξkα in the sub-critical regime.

Remark 4.1.3. Let us notice that in the Brownian setting, the position X and the velocity
V become independent in large time. However, this is false in the stable case. Indeed,

147
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

the limit is a rotationally invariant stable process on R2 , which cannot have independent
coordinates.

Remark 4.1.4. The rate of convergence of the position process is the same as for the ve-
locity process. This is not true in the free potential case, i.e. when U = 0 (see [GL21b]). In-
deed, our rate of convergence for X is smaller. For example in the Brownian super-critical
regime, the variance of Xt heuristically behaves as 2t in large time, while in [GL21a], it
3
behaves as t3 . This is quite natural because the quadratic potential tends to confine the
particle through a spring force, so the particle spreads out more slowly than without po-
tential.

Remark 4.1.5. The condition α ∈ (1, 2) and the symmetry of ν are only required to
ensure the well-posedness of (SKE) when γ < 1. Nevertheless, proofs of the moment
estimates and of the asymptotic behavior can be written for a non-symmetric α-stable
Lévy driving process with α ∈ (0, 2).

4.2 Existence up to explosion

Theorem 4.2.1. Assume that α > 1. The system (SKE) admits a unique weak solution
if γ ∈ (0, 1]. If γ > 1, there exists a unique strong solution defined up to its explosion time
τ∞ .

Proof. In the case γ > 1, the coefficients of the SDE (4.2) satisfied by Z = (X, V ) are
locally Lipschitz continuous with respect to the space variable, locally uniformly in time.
So we can apply Lemma 115 p. 78 in [Sit05] to get the pathwise uniqueness. The drift
coefficient is continuous with respect to its two variables, so it is a locally bounded and
measurable function. By a standard localization argument, using Theorem 9.1 p. 231 in
[IW81], since the drift coefficient is locally Lipschitz, there is a unique solution defined up
to explosion.
Assume now that γ ≤ 1. We check that we can use Theorem 1 in [MM21]. Using the same
notations, we have  
0 −1
Ae =  ,
1 0

and for any (t, x1 , x2 ) ∈ [t0 , +∞) × R2 , F1 (t, x1 , x2 ) = sgn(x1 )|x1 |γ t−β , F2 (t, x1 , x2 ) = 0
and σ(t, x1 , x2 ) = 1. Assumptions (UE) and (ND) are clearly satisfied. Since F2 does not
depend on x1 and since [A] 2,1 = 1 is different from 0, we deduce that Assumption (H)
e

148
 4.3. Moment estimates and non-explosion

is satisfied. We easily check that Theorem 1 in [MM21] can be applied with β1 = γ, and
β2 = 1.

Remark 4.2.2. For α ∈ (0, 2), employing the technique of Picard iteration and the in-
terlacing procedure, one can deduce that (4.2) has a unique solution in the linear setting
γ = 1 (see [App09, p. 375]).

4.3 Moment estimates and non-explosion

Let Z be the unique solution up to explosion time to (4.2). As in the continuous
setting, define, for all r ≥ 0, the stopping time

τr := inf{t ≥ t0 , kZt k ≥ r}.

Set τ∞ := limr→+∞ τr the explosion time of Z. For the sake of simplicity, since there is no
jump on the position component, for z ∈ R, we shall write Zs− + z for (Xs , Vs− + z) in
the following.
We adapt the proof of [GL21b] to two-dimensional processes.

Proposition 4.3.1. Pick α ∈ (0, 1). For any γ, β, the explosion time τ∞ is a.s. infinite
and for all κ ∈ [0, α), there exists a constant Cκ,t0 such that, we have

κ
∀t ≥ t0 , E [kZt kκ ] ≤ Cκ,t0 t α . (4.3)

Remark 4.3.2. Note that, as in the Brownian case, the moment estimates obtained for
the position process X is a priori smaller in our case than in the free potential case
[GL21b]. It is explained by the confining effect of the quadratic potential.

Proof. Fix t ≥ t0 and κ ∈ [0, α). Since α < 1, the symmetric stable process can be written
as Z tZ X
Lt = zN (ds, dz) = ∆Ls .
0 R∗ s≤t
q
Pick the sequence of C 2 -functions fn : (x, v) 7→ x2 + v 2 + n1 , which converges uniformly
to (x, v) 7→ k(x, v)k on R2 . Then, for all n ≥ 1 and r ≥ 0, we apply Itô’s formula (see
Theorem 33 p. 81 in [Pro05]). We use the fact that for all y ∈ R2 , y · Ay = 0, and observe

149
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

that for any s ≥ t0 and (x, v) ∈ R2 , (x, v) · F (s, v) = |v|γ+1 s−β ≥ 0. Namely, we get
X
fn (Zt∧τr ) ≤ fn (z0 ) + (fn (Zs ) − fn (Zs− )).
s≤t∧τr

Since sup{k∇fn (y)k, y ∈ R2 } ≤ 1, we deduce, from Taylor-Lagrange’s inequality that for

all s ≤ t ∧ τr , |fn (Zs ) − fn (Zs− )| ≤ k∆Zs k = |∆Ls |. It follows that
X
kZt∧τr k ≤ fn (Zt∧τr ) ≤ fn (z0 ) + |∆Ls | .
s≤t∧τr

Furthermore, since κ < α < 1, we have

 κ 

E [kZt∧τr kκ ] ≤ E [fn (z0 )κ ] + E  |∆Ls |  .

X

s≤t

Notice that the process L+ |∆Ls | is an α-stable process. Then, since κ < α,
P
t := s≤t
letting n → +∞, we obtain
h κi κ
E [kZt∧τr kκ ] ≤ kz0 kκ + E L+
t ≤ Ct0 ,κ t α .

Thanks to Lemma B.0.1, we can conclude that the explosion time of Z is a.s. infinite and
Fatou’s lemma yields (4.3).

Proposition 4.3.3. Pick α ∈ [1, 2). For any γ ≥ 0 and β ≥ 0, the explosion time τ∞ is
a.s. infinite and for κ ∈ (0, α), there exists Cκ,t0 such that

κ
∀t ≥ t0 , E [kZt kκ ] ≤ Cκ,t0 t α . (4.4)

Proof. The key idea is to slice the small and big jumps in a non-homogeneous way with
1
respect to the characteristic scale of an α-stable process ξ 7→ ξ α .
Pick ξ ≥ t0 . The α-stable symmetric Lévy driving process can be written as
Z tZ Z tZ
Lt − Lt0 = 1
f(ds, dz) +
zN 1 zN (ds, dz).
t0 |z|≤ξ α t0 |z|>ξ α

Step 1. We first apply Itô’s formula (see Theorem 4.47 p. 251 in [App09]) and estimate
the expectation of each term for κ ≤ 1, in order to get (4.4).
Fix η > 0 to be chosen latter and define the C 2 -function f : (x, v) 7→ (η + x2 + v 2 )κ/2 . We

150
 4.3. Moment estimates and non-explosion

use the fact that for all y ∈ R2 , y · Ay = 0, and observe that for any s ≥ t0 and (x, v) ∈ R2 ,
(x, v)T · F (s, v) = |v|γ+1 s−β ≥ 0. For all t ≥ t0 , by Itô’s formula, we have

f (Zt∧τr ) ≤ f (V0 ) + Mt + Rt + St ,

where Z tZ
Mt := 1 1{s≤τr } (f (Zs− + z) − f (Zs− ))N
f(ds, dz),
t0 0<|z|<ξ α

Z tZ
Rt := 1 1{s≤τr } (f (Zs− + z) − f (Zs− ))N (ds, dz), (4.5)
t0 |z|≥ξ α

Z tZ
St := 1 1{s≤τr } [f (Zs− + z) − f (Zs− ) − df (Zs− ).z] ν(dz) ds. (4.6)
t0 0<|z|<ξ α

Moreover, remark that for all k > α,

Z
2a k −1
1 |z|k ν(dz) = ξα , (4.7)
0<|z|<ξ α k−α

and for all k < α, Z

2a k −1
1 |z|k ν(dz) = ξα . (4.8)
|z|≥ξ α α−k
We estimate expectations of M , R and S.
To that end, we first show that the local martingale (Mt )t≥t0 is a martingale. Fix q ≥ 2
and r ≥ 0. Moreover, we set
Z tZ
It (q) := 1 1{s≤τr } |f (Zs− + z) − f (Zs− )|q ν(dz) ds.
t0 0<|z|<ξ α

1
Thanks to Taylor-Lagrange inequality, for all k(x, v)k ≤ r and |z| ≤ ξ α ,

|f (x, v + z) − f (x, v)| ≤ sup{k∇f (y)k , kyk ∈ [−r − ξ /α , r + ξ /α ]} |z| ≤ Cr,ξ,κ |z| ,
1 1

so we have Z tZ
It (q) ≤ Cr,ξ,κ 1 1{s≤τr } |z|q ν(dz) ds.
t0 0<|z|<ξ α

Hence, it is a finite quantity, since q ≥ 2 and (4.7) holds. Therefore, for q ≥ 2, by Kunita’s

151
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

inequality (see Theorem 4.4.23 p. 265 in [App09]), there exists Dq > 0 such that
" #
q
 h q i 
E sup |Ms | ≤ Dq E It (2) 2 + E [It (q)] < +∞.
t0 ≤s≤t

Hence, by Theorem 51 p. 38 in [Pro05], M is a martingale.

We estimate now the finite variation part S defined in (4.6). Note that for all (x, v) ∈ R2 ,
the Hessian matrix of f is given by

x2 xv
 
1 + (κ − 2) 2 2
(κ − 2)
x +v +η x + v2 + η 
2
κ
 
Hess(f )(x, v) = κ(x2 + v 2 + η) 2 −1  2 .
xv v

(κ − 2) 2 −
 
2
1 + (κ 2) 2 2
x +v +η x +v +η
κ
Its matrix norm is bounded by Cκ η 2 −1 .
1
Assume that |z| < ξ α . Using Taylor- Lagrange’s inequality and injecting (4.7) we get the
almost sure following bound, for all s ≥ t0 ,
Z
κ 2a 2 −1
1 (f (Zs− + z) − f (Zs− ) − ∇f (Zs− ) · z) ν(dz) ≤ Cκ η 2 −1 ξα . (4.9)
0<|z|<ξ α 2−α

It remains to study the Poisson integral R defined in (4.5). Pick κ ≤ 1, by Hölder property
of power functions and (4.8), we deduce that
Z
κ 2a −1 2a κ −1
1 |f (Zs− + z) − f (Zs− )| ν(dz) ≤ η 2 ξ + ξα . (4.10)
|z|≥ξ α α α−κ

Moment estimate of the Poisson integral follows from Theorem 2.3.7 p. 106 in [App09].
Gathering (4.10) and (4.9), we obtain

2a 2a κ 2a 2
 
κ
E [kZt∧τr kκ ] ≤ E [f (Zt∧τr )] ≤ E [f (Zt0 )]+tξ −1 η κ/2 + ξ α + Cκ η 2 −1 ξα .
α α−κ 2−α
2
Choosing η = t α and ξ = t, we get

2a 2a 2a
 
κ κ κ
E [kZt∧τr k ] ≤ E [f (Zt0 )] + t α + + Cκ ≤ Cκ,t0 t α .
α α−κ 2−α

Thanks to Lemma B.0.1, we can conclude that the explosion time of Z is a.s. infinite, and

152
 4.4. Asymptotic behavior of the solution

letting r → +∞ with Fatou’s lemma, for all κ ∈ [0, 1],

κ
E [kZt kκ ] ≤ Cκ,t0 t α . (4.11)

Step 2. Pick κ ∈ (1, α). We estimate R in another way, using again Theorem 2.3.7 p. 106
in [App09].
By the Hölder property of power function and (4.8), we get
Z Z κ

1 |f (Zs− + z) − f (Zs− )| ν(dz) ≤ 1 2zVs− + z 2 2

ν(dz)
|z|≥ξ α |z|≥ξ α
! (4.12)
2a κ −1 κ 2a κ
−1
≤ Cκ ξ α + |Vs− | 2 κξ
2α .
α−κ α− 2

Gathering (4.9) and (4.12), one has

2a κ −1 2a 2 −1
 
κ κ
E [kZt∧τr k ] ≤ E [f (Zt0 )] + t Cκ ξ α + Cκ η 2 −1 ξα
α−κ 2−α
2a κ
−1
Z t h
κi
+ Cκ ξ 2α E |V s | 2 ds.
α − κ2 t0

2
Injecting (4.11) applied with κ2 , choosing η = t α and ξ = t, we get

κ
E [kZt∧τr kκ ] ≤ Cκ,t0 ,α t α .

The conclusion of the proof follows, letting r → +∞.

4.4 Asymptotic behavior of the solution

We gather in this section the proof of Theorem 4.1.1. The strategy is to prove the
convergence of the f.d.d. of the process Y (ε) , and then its tightness both in the super-
critical and critical regimes. We first prove the tightness when αq ≥ 1.
n 1
o
Lemma 4.4.1. Assume that αq ≥ 1, then the family (ε α Yt/ε )t≥εt0 , ε > 0 is tight on
every compact interval [m, M ], for 0 < m ≤ M .

Proof. We check the Aldous’s tightness criterion stated in Theorem 16.10 p. 178 in [Bil99].
Let a, η, T be positive reals. Let τ be a discrete stopping time with finite range T ,

153
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

bounded by T and fix δ > 0 and ε > 0 small enough to be chosen latter. Define, for
t ≥ εt0 , the Wiener-Lévy integral appearing in (SDEY )
 
(ε) 1
Z t/ε
1
Z t/ε
− sin(s)
Mt := ε α e−sA Γ dSs = ε α  dLs . (4.13)
t0 t0 cos(s)

We use Jensen’s inequality to get, for r = α2 ,

" Z
(τ +δ)/ε r# " Z (τ +δ)/ε #r
h
(ε) ri 1 1 γ
−uA −β
E Yτ +δ − Yτ(ε) ≤E ε e
α Γ dSu +E ε α |Vu | u du .
τ /ε τ /ε

The stopping time has a finite range T , hence, we can write

" Z
(τ +δ)/ε r# " " Z
(τ +δ)/ε r ##
1 1
−uA −uA
E ε e
α Γ dSu =E E ε e α Γ dSu τ
τ /ε τ /ε
" " r ##
X 1{τ =τi } Z (τi +δ)/ε
1
−uA
=E E 1{τ =τi } εα e Γ dSu
τi ∈τ P(τ = τi ) τi /ε
" " Z r ##
X 1{τ =τi } (τi +δ)/ε 1
−uA
≤E E ε e α Γ dSu .
τi ∈τ P(τ = τi ) τi /ε

Besides, we can use the self-similarity of S, Lemma 5.2 in [DF13], and the fact that etA
is a rotation matrix for any t ∈ R, to compute for each τi ∈ T and δ small enough
" Z
(τi +δ)/ε r# " Z
τi +δ r#
1
−uA − uε A r
E ε e
α Γ dSu =E e Γ dSu ≤ Cr δ α .
τi /ε τi

Since τ ∈ [m, M ] a.s., the last term can be handled as in Section 5.2 in [GL21b] using
moment estimates of V (see Propositions 4.3.1 and 4.3.3) to have
" Z (τ +δ)/ε #r
1 γ γ+α−1
−β
E ε α K |Vu | u du ≤ Cm,M εrβ−r α .
τ /ε

Since η > 0 and by Markov’s inequality, we obtain for δ and ε small enough
r γ+α−1

(ε) Cr δ α + Cm,M εrβ−r
 α
P Yτ +δ − Yτ(ε) ≥a ≤ ≤ η.
ar

Moreover, by Markov’s inequality and the moment estimates again, we deduce that for

154
 4.4. Asymptotic behavior of the solution

all t ∈ [m, M ],

(ε) r
h i
r

(ε)
 E Yt Ct α
lim lim sup P Yt ≥ a ≤ lim lim sup ≤ lim = 0.
a→+∞ ε→0 a→+∞ ε→0 ar a→+∞ ar

By Corollary and Theorem 16.8 p. 175 in [Bil99], this concludes the proof of the tightness
on every compact interval of (0, +∞).

We will now prove the convergence of the f.d.d. of Y (ε) . Thanks to the previous lemma,
this will yield the weak convergence on every compact set (see Theorem 13.1 p. 139
in [Bil99]) in the super-critical and critical regimes. The convergence in distribution on
the whole space D will follow from Theorem 16.7 p. 174 in [Bil99], since all processes
considered are càdlàg.

4.4.1 Convergence of the f.d.d. in the super-critical regime

(ε) 1
Assume that αq > 1. Recall that (Yt )t≥εt0 := (ε α Yt/ε )t≥εt0 .

Proof of Theorem 4.1.1 (i).

Step 1. We first prove the convergence of the f.d.d. of the Wiener-Lévy integral appearing
in (SDEY ).
Recall that the local martingale M (ε) was defined in (4.13).
(ε) (ε)
Step 1a. We begin with the convergence in distribution of Ms,t := Mt − Ms(ε) , for
εt0 ≤ s ≤ t.
(ε) (ε)
To this end, we study the characteristic function φs,t of Ms,t . Let us recall that ψ denotes
the characteristic exponent of L, and is given, for all ξ ∈ R, by

ψ(ξ) = −a|ξ|α .

The characteristic function of the Wiener-Lévy integral can be computed as p. 105 in

155
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

[Sat99], hence one has, for all ξ := (u, v) ∈ R2 ,

" Z t/ε Z t/ε #!
(ε) 1 1
φs,t (ξ) = E exp −iuε α sin(y) dLy + ivε α cos(y) dLy
s/ε s/ε
" Z t/ε #!
1
= E exp iε α (−u sin(y) + v cos(y)) dLy
s/ε
Z t/ε !
1
 
= exp ψ ε [−u sin(y) + v cos(y)] dy
α
s/ε
Z t/ε !
α
= exp −aε | − u sin(y) + v cos(y)| dy .
s/ε

(ε)
Using Lemma B.0.2, we deduce that φs,t (ξ) converges, as ε → 0, to

1 Z 2π
 
exp −a(t − s) |−u sin(y) + v cos(y)|α dy .
2π 0

Step 1b. We now compute explicitly the scale parameter of the stable limiting process.
We denote by λ the uniform probability distribution on the circle S1 . Thanks to a change
ξ
of variable and the symmetry of λ, setting ω := kξk for ξ = (u, v) ∈ R2 \ {0}, we have

1 Z 2π α
Z
|−u sin(y) + v cos(y)| dy = |ξ · λ|α dλ
2π 0 S1
Z
= kξkα |ω · λ|α dλ.
S1

Since λ is rotationally invariant, we deduce that S1 |ω · λ|α dλ does not depend on ω ∈ S1 .

R

Taking ω = (1, 0)T , we set

a Z 2π
Ce := |cos(x)|α dx. (4.14)
2π 0
We have thus proved that, for any ξ ∈ R2 ,
 
(ε) α
φs,t (ξ) −→ exp −(t − s)Ckξk
e .
ε→0

Thus, the following convergence in distribution holds

(ε) (ε)
Ms,t = Mt − Ms(ε) =⇒ Lt−s . (4.15)
ε→0

156
 4.4. Asymptotic behavior of the solution

Following the same lines, we show that, for any t > 0,

(ε)
Mt =⇒ Lt . (4.16)
ε→0

Step 1c. We now prove the convergence in f.d.d. of M (ε) to L, as ε tends to 0.

(ε)
Let us fix 0 < t1 ≤ t2 ≤ · · · ≤ td . Note that (Mt )t≥εt0 is a càdlàg process with inde-
pendent increments, since the integrands in its definition are deterministic and because
(ε) (ε) (ε)
L is a Lévy process. Thus, the random variables (Mt1 , Mt1 ,t2 , . . . , Mtd−1 ,td ) are mutually
independent. We deduce from the convergence results established in (4.15) and (4.16),
and the fact that L has stationary and independent increments that

(ε) (ε) (ε)

(Mt1 , Mt1 ,t2 , . . . , Mtd−1 ,td ) =⇒ (Lt1 , Lt2 − Lt1 , . . . , Ltd − Ltd−1 ).
ε→0

The continuous mapping theorem yields the convergence in f.d.d. of M (ε) to L.

Step 2. Pick T > 0. We prove that
" #
(ε) (ε)
E sup Yt − Mt −→ 0.
εt0 ≤t≤T ε→0

We have Z T /ε
(ε) (ε) 1 1
sup Yt − Mt ≤ ε kz0 k + ε
α α e−sA F (s, Vs ) ds.
εt0 ≤t≤T t0

We use moment estimates (Propositions 4.3.1 and 4.3.3) to get

" Z T /ε # " Z T /ε #
1 1
−sA
E ε α e F (s, Vs ) ds = E ε α kF (s, Vs )k ds
t0 t0
" Z T /ε #
1 γ −β
≤E ε α |Vs | s ds
t0
Z T /ε
1 γ
≤ ε α Cκ,t0 s α −β ds
t0
γ
γ+α−1 γ 1 −β+1
≤ Cκ,t0 (εβ− α T α −β+1 − ε α t0α ).

γ+α−1 1
Hence, setting r := min(β − α
, α ), which is positive by assumption, we get
" #
(ε) (ε)
E sup Yt − Mt = O (εr ). (4.17)
εt0 ≤t≤T ε→0

The conclusion follows from Theorem 3.1 p. 27 in [Bil99].

157
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

4.4.2 Convergence of the f.d.d. in the critical and sub-critical

regime
 i
1
In this section, we consider the linear case, i.e. γ = 1 and we assume that β ∈ 2
,1 .
Recall that
 
(ε)
(Yt )t≥εt0 = εq Yt/ε .
t≥εt0

Proof of Theorem 4.1.1 (ii) and (iii). The proof follows the same lines as in the Brownian
setting. Recall that leaving out the Brownian term, the underlying ODE is the following

x0 (t)
x00 (t) + + x(t) = 0, t ≥ t0 . (4.18)
tβ

We pick again the basis of solutions given by Lemma A.0.3, and we still denote by R its
resolvent matrix and by f its rate of decrease. Recall that it is given by

 √1

if β = 1,
t
∀t > 0, f (t) :=   (4.19)
t1−β
exp − 2(1−β) else.


We set, for all t ≥ εt0 ,

Z t/ε
f (ε)
M q
:= ε f (t/ε) Rs−1 Γ dSs . (4.20)
t
t0

Keeping the same notations as in the Brownian case, we decompose (Yt )t≥t0 = (e−tA Zt )t≥t0
into
εq Yt/ε = εq f (t/ε)Φt/ε Rt−1 f (ε) .
Z0 + Φt/ε M
0 t

Reasoning as in the Brownian case, it remains to study the convergence of Mf (ε) since

the first term converges towards 0. Using the expression of the Wronskian obtained in
Lemma A.0.3, we obtain, for all t ≥ εt0 ,
 
(ε)
Z t/ε
−y2 (u)
M
f
t = εq f (t/ε) f (u)−2  dLu .
t0 y1 (u)

f (ε) , M (ε)
Let us fix 0 < s < t. We study the convergence in distribution of the couple (M s
f )
t
when ε tends to 0. The convergence in distribution of a general d-dimensional distribution
f (ε) , . . . , M
(M f (ε) ) relies on the same computations.
t1 td
Let us fix (ξ1 , ξ2 ) ∈ R2 × R2 . Using that L has independent increments, the characteristic

158
 4.4. Asymptotic behavior of the solution

(ε) f (ε) ) is given by

f (ε) , M
function φes,t of (Ms t

    
(ε)
Z s/ε
−y2 (u)
φes,t (ξ1 , ξ2 ) = E exp iεq f (s/ε) ξ1 · f (u)−2  dLu
t0 y1 (u)
  
Z t/ε
−y2 (u)
+f (t/ε) ξ2 · f (u)−2  dLu 
t0 y1 (u)
    
q
Z s/ε
−2 −y2 (u)
= E exp iε (f (s/ε) ξ1 + f (t/ε) ξ2 ) · f (u)  dLu 
t0 y1 (u)
    
q
Z t/ε
−2 −y2 (u)
× E exp iε f (t/ε) ξ2 · f (u)  dLu  .
s/ε y1 (u)

Let us recall that the characteristic exponent of L is given, for all ξ ∈ R, by

ψ(ξ) = −a|ξ|α .

The characteristic function of the Wiener-Lévy integral can be computed as p. 105 in

[Sat99], hence one has
  α 
(ε)
Z s/ε
−y2 (u)
φes,t (ξ1 , ξ2 ) = exp −aεβ f (u)−2α (f (s/ε) ξ1 + f (t/ε) ξ2 ) ·  du
t0 y1 (u)
  α 
Z t/ε
−y2 (u)
× exp −aεβ f (u)−2α f (t/ε) ξ2 ·  du .
s/ε y1 (u)

Using the asymptotic expansion of the resolvent matrix (Lemma A.0.3), we can write, for
any u ≥ t0 ,     

−y 2 (u) = f (u) 
− sin(u)  + g(u) ,
y1 (u) cos(u)
where g : [t0 , +∞) → R2 is a function satisfying for all u ≥ t0 ,

|g(u)| ≤ Cu1−2β .

159
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

We set
   α 
(ε)
Z s/ε
− sin(u)
K1 := exp −aεβ f (u)−α (f (s/ε) ξ1 + f (t/ε) ξ2 ) ·  + g(u) du
t0 cos(u)
(4.21)
and    α 
Z t/ε
t − sin(u)
 
(ε)
K2 := exp −aεβ f (u)−α f ξ2 ·  + g(u) du .
s/ε ε cos(u)
We thus obtain
(ε) (ε) (ε)
φes,t (ξ1 , ξ2 ) = K1 × K2 , (4.22)

Step 1. We start by justifying that we can omit g to study the limit when ε → 0. More
precisely, we prove that, for all function ζ : R → R2 such that kζ(ε)k f (s/ε)−1 = O (1),
ε→0

  α   α
Z s/ε
− sin(u) − sin(u)
R(ε) := εβ f (u)−α ζ(ε) ·  + g(u) − ζ(ε) ·  du −→ 0.
t0 cos(u) cos(u) ε→0

(4.23)
Thanks to the mean value theorem applied to | · |α (for α ≥ 1) or by Hölder property (for
α < 1), and the domination of g, we obtain that, for some constant C > 0,
Z s/ε
R(ε) ≤ Cεβ kζ(ε)kα f (u)−α u1−2β du = O (kζ(ε)kα f (s/ε)−α ε2β−1 ),
t0 ε→0

where the last equality follows from Lemma B.0.3. This proves (4.23) since β > 21 .
(ε)
Step 2. We focus on the first term K1 defined in (4.21). Since f is decreasing, notice
that
ζ(ε) := f (s/ε) ξ1 + f (t/ε) ξ2 = O (f (s/ε)).
ε→0

Then we have to study the convergence of I (ε) defined by

 α
Z s/ε
− sin(u)
I (ε) := aεβ f (u)−α (f (s/ε) ξ1 + f (t/ε) ξ2 ) ·  du.
t0 cos(u)

Its limit differs according to the value of β.

Step 2a. Assume first that β = 1. Then, using the expression of f (see (4.19)),
!  α
1+ α
Z s/ε
ξ ξ − sin(u)
I (ε) = aε 2 f (u)−α √1 + √2 ·  du.
t0 s t cos(u)

160
 4.4. Asymptotic behavior of the solution

We proved in Step 2b of the super-critical regime that there exists a constant Ce > 0 given
in (4.14) such that, for all ζ ∈ R2 ,
α 
a Z 2π − sin(u) α
ζ ·  du = Ckζk
e . (4.24)
2π 0 cos(u)

Using Lemma B.0.2, we can compute the following asymptotic expansion

α Z s/ε Z s/ε !
1+ α ξ ξ 1+ α
I (ε)
=ε 2 Ce √1 + √2 f (u) −α
du + o ε 2 f (u) −α
du .
s t t0 ε→0 t0

Therefore, it follows from Lemma B.0.3 that

−1 α !
α ξ ξ

1+ α
(ε)
K1 −→ exp −Ce 1+ √1 + √2 s 2 .
ε→0 2 s t
 
1
Step 2b. Let us consider now β ∈ 2
,1 . Let us notice that I (ε) can be decomposed into
the sum
(ε) (ε)
I (ε) = I1 + I2 (4.25)

of the two following terms

 α
Z s/ε
s − sin(u)
 
(ε)
I1 := aεβ f (u)−α f ξ1 ·  du
t0 ε cos(u)

and
  α
Z s/ε
s t − sin(u)
     
(ε)
I2 := aεβ f (u)−α  f ξ1 + f ξ2 · 
t0 ε ε cos(u)
  α
s − sin(u) 
 
− f ξ1 ·  du.
ε cos(u)

Using again either the mean value theorem or the Hölder property, and Lemma B.0.3, we
get since β < 1 that for some constant C > 0,
Z s/ε  
(ε)
|I2 | ≤ Cεβ f (s/ε)α−1 f (t/ε) f (u)−α du = O f (t/ε)f (s/ε)−1 = o (1). (4.26)
t0 ε→0 ε→0

161
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

(ε)
Using Lemma B.0.2, we can compute the following asymptotic expansion of I1
 α
− sin(u)
 α Z s/ε
(ε) s
I1 = aεβ f f (u)−α ξ1 ·  du
ε t0 cos(u)
  α  ! 
−
 α
s 1 Z 2π
sin(u) Z s/ε Z s/ε
= aεβ f  ξ1 ·   du f (u)−α du + o f (u)−α du  .
ε 2π 0 cos(u) t0 ε→0 t0

Thanks to (4.24) and the asymptotic expansion’s results given in Lemma B.0.3, there
exists an explicit constant kβ,α given in Lemma B.0.3, such that

(ε) e β kξ kα .
I1 −→ kβ,α Cs 1 (4.27)
ε→0

(ε)
Combining (4.23), (4.25), (4.26) and (4.27), we have proved that K1 , defined in (4.21),
converges as ε → 0 towards
 
exp −kβ,α Cse β kξ kα .
1

(ε)
Step 3. It remains to deal with the limit of K2 . Notice that

ζ(ε) := f (t/ε) ξ2 = O (f (t/ε)) = O (f (s/ε)) .

ε→0 ε→0

Hence, thanks to Step 1, we are reduced to study, for r ∈ {s, t},

 α
 α Z r/ε
t − sin(u)
Jr(ε) := aεβ f f (u)−α ξ2 ·  du. (4.28)
ε t0 cos(u)

Asymptotic expansion’s results (Lemmas B.0.2 and B.0.3) and (4.24) yield
Z r/ε  
Jr(ε) = Ce kξ 2k
α β
ε f (t/ε) α
f (u)−α du + o f (t/ε)α f (r/ε)−α
t0 ε→0
 
−α
= Ce kξ 2k
α β
kβ,α r f (t/ε) f (r/ε)α
+ o f (t/ε)α f (r/ε)−α .
ε→0

Hence,
(ε)
Jt −→ Ce kξ2 kα kβ,α tβ (4.29)
ε→0

and  α
α β s 2
Js(ε) −→ Ce kξ 2k kβ,α s 1{β=1} . (4.30)
ε→0 t

162
 4.5. Proof of Corollary 4.1.2

Since
 
(ε) (ε)
K2 = exp −Jt + Js(ε) ,

we thus obtain that, for all 0 < s ≤ t,

    
e β kξ kα exp −k Ct
exp −kβ,α Cs
 e β kξ kα if β < 1,
(ε) 1 β,α 2
φes,t (ξ1 , ξ2 ) −→  
ξ ξ α α α α
  α 
ε→0 
exp −k C 1+ s 2
e √ +√
β,α
1 2
s
s 2 + kξ k t − kξ
t 2 2k t
s if β = 1.

 R √ 
Step 4. We can compute the characteristic function of the process √1t 0t s dLs in
t>0
the same manner, and thus recognize the limiting process in the critical regime.

Remark 4.4.2. As in the Brownian setting, if β = 0, the resolvent matrix is explicit

 
and following the same lines, we can prove that Zt/ε converges in f.d.d. towards the
t≥εt0
product of the measure µ, whose characteristic function is given by
 Z +∞ 
−αu α
ξ 7→ exp −a e kξ · h(u)k du ,
0

h being an explicit periodic function depending on the resolvent matrix.

4.5 Proof of Corollary 4.1.2

Proof of Corollary 4.1.2. We start by proving the convergence in distribution of v1/T ZT .
Reasoning as in the Brownian setting, it follows from Theorem 4.1.1 that v1/T YT converges.
The conclusion is a consequence of Lemma B.0.4, noting that the limiting distribution is
invariant under rotations thanks to the expression of its characteristic function.
Let us now prove that the rescaled process Z (ε) does not converge in distribution. We state
the proof in the super-critical regime. Assume by contradiction that it is the case. Rea-
soning as in the Brownian case, we prove that this implies the convergence in distribution
of the process I (ε) defined, for t ≥ εt0 , by
Z t/ε
t
 
(ε) 1
It := ε α sin − u dLu .
t0 ε
(ε)
In particular, for s < t, the random variable It − Is(ε) shall converge in distribution.
(ε)
Let us denote by φ(ε) the characteristic function of It −Is(ε) , which is supposed to converge

163
Chapter 4 – Behavior of the non-autonomous system driven by a Lévy process

on R. Using that L has independent increments, and setting

" Z t/ε !#
t
 
1
φ(ε),1 := E exp ε α sin − u dLu ,
s/ε ε
" Z s/ε !#
t s
   
1
(ε),2
φ := E exp ε α sin − u − sin − u dLu ,
t0 ε ε
we have φ(ε) (1) = φ(ε),1 φ(ε),2 .
Recall that ψ defined in (4.1), denotes the characteristic exponent of L. Using a change
of variables, we have in particular
Z t/ε ! Z t/ε α !
t t
   
1
(ε),1
φ = exp ψ ε sin −u α du = exp −aε sin −u du
s/ε ε s/ε ε
Z (t−s)/ε !
α
= exp −aε |sin (u)| du .
0

Lemma B.0.2 ensures that φ(ε),1 has a limit when ε converges to 0. Similarly, we obtain
Z s/ε !
t s
    
1
(ε),2
φ = exp ψ ε sin α − u − sin − u du
t0 ε ε
Z s/ε α !
t s
  
= exp −aε sin − u − sin −u du
t0 ε ε
!
t − s α Z s/ε
α
t+s
  
α
= exp −a2 sin ε cos −u du
2ε t0 2ε
t+s
!
t − s α Z 2ε −t0
 
α α
= exp −a2 sin ε t−s |cos (u)| du . (4.31)
2ε 2ε

The change of variables u = v + π yields, for all ε > 0,

Z 2π
t−s
   
|cos (u)|α sgn sin cos (u) du = 0.
0 2ε

Thus, Lemma B.0.2 ensures that

t+s
Z
2ε
−t0 s Z 2π
ε |cos (u)|α du −→ | cos(u)|α du.
t−s
2ε
ε→0 2π 0

Coming back to (4.31), we see that φ(ε),2 does not converge when ε tends to 0. This is a
contradiction.

164
 Appendix A

S TUDY OF THE DETERMINISTIC

UNDERLYING ODE

The deterministic ODE behind the system is the following

x0 (t)
x00 (t) + + x(t) = 0, t ≥ t0 . (A.1)
tβ

The solutions form a vector space of dimension 2. Let us take two solutions y1 and y2
which are linearly independent. Then, we introduce the fundamental system of solutions
(resolvent matrix) R to (A.1) defined, for t ≥ t0 , by
 
y1 (t) y2 (t)
Rt =  0 .
y1 (t) y20 (t)

It satisfies, for all t ≥ t0 ,  

0 1 
Rt0 =  Rt .
−1 − t1β
We recall that the Wronskian w is defined, for all t ≥ t0 , by

w(t) = y1 (t)y20 (t) − y10 (t)y2 (t).

Let us finally set, for t > 0,


 √1

if β = 1,
t
f (t) :=  1−β
 (A.2)
t
exp − 2(1−β) else.


 i
1
Lemma A.0.1. Pick β ∈ 2
,1 and consider a solution y to (A.1). Then, there exist

165
Chapter A – Study of the deterministic underlying ODE

a ∈ R and φ ∈ [0, 2π) such that

 
y(t) = af (t) cos(t + φ) + O f (t)t1−2β
t→∞

and
 
y 0 (t) = −af (t) sin(t + φ) + O f (t)t1−2β .
t→∞

Proof. Let us set, for t ≥ t0 , u(t) = f (t)−1 y(t). We easily check that u satisfies

u00 (t) + u(t) [1 + h(t)] = 0,

00 0
where h(t) := ff (t)
(t)
+ ff(t)t
(t)
β = O (t−2β ). Following the proof of the method of variation
t→+∞
of parameters, there exists a0 , b0 ∈ R such that, for any t ≥ t0 ,
Z t
u(t) = a0 cos(t) + b0 sin(t) − u(s)h(s) sin(t − s) ds.
t0

Using that h ∈ L1 ((t0 , +∞)) since β > 12 , we obtain by Grönwall’s lemma that the function
u is bounded on [t0 , +∞). We deduce that both the functions s 7→ u(s)h(s) cos(s) and
s 7→ u(s)h(s) sin(s) belong to L1 ((t0 , +∞)). Thus, up to changing the constants a0 and
b0 , one has, for all t ≥ t0 ,
Z ∞ Z ∞
u(t) = a0 cos(t) + b0 sin(t) + sin(t) u(s)h(s) cos(s) ds − cos(t) u(s)h(s) sin(s) ds.
t t
(A.3)
It follows from the fact that u is bounded that
Z ∞ !
ds
u(t) = a0 cos(t) + b0 sin(t) + O .
t→+∞ t s2β

Thus, there exist a ∈ R and φ ∈ [0, 2π) such that

u(t) = a cos(t + φ) + O (t1−2β ).

t→+∞

This proves the asymptotic expansion of y.

Differentiating (A.3) and using that h(t) = O (t−2β ), we prove that
t→+∞

u0 (t) = −a sin(t + φ) + O (t1−2β ).

t→+∞

166
Since u is bounded and f 0 (t) = O (f (t)t−β ), we finally obtain that
t→+∞

y 0 (t) = f 0 (t)u(t) + f (t)u0 (t) = f (t)u0 (t) + O (f (t)t−β ).

t→+∞

This concludes the proof of the asymptotic expansion of y 0 since β ≥ 2β − 1.

Remark A.0.2. Note that if β = 1, the Bessel functions of the first kind J0 and of the
second kind Y0 form a basis of solutions. Their asymptotic expansions can be found in
[Wat44, Chap VII].

Lemma A.0.3. There exists a basis of solutions y1 and y2 to (A.1) such that the resolvent
matrix R satisfies
 
cos(t) sin(t)     
Rt = f (t)  + O f (t)t1−2β = f (t)etA + O f (t)t1−2β .
− sin(t) cos(t) t→∞ t→∞

Moreover, its Wronskian w is given, for any t ≥ t0 , by w(t) = f (t)2 .

Proof. It is well-known that the Wronskian satisfies, for all t ≥ t0 ,

1
w0 (t) = − w(t).
tβ

Thus, there exists w0 ∈ R \ {0} such that, for all t ≥ t0 , w(t) = w0 f (t)2 . Moreover, thanks
to Lemma A.0.1, for i ∈ {1, 2}, there exist ai ∈ R and φi ∈ [0, 2π) such that
 
yi (t) = ai f (t) cos(t + φi ) + O f (t)t1−2β
t→∞

and
 
yi0 (t) = −ai f (t) sin(t + φi ) + O f (t)t1−2β .
t→∞

As a consequence,
 
w(t) = −a1 a2 f (t)2 sin(φ2 − φ1 ) + O f (t)2 t2−4β .
t→∞

But since w(t) = w0 f (t)2 , it implies that ai 6= 0 and φ2 6≡ φ1 [π].

Up to dividing by ai , we can assume that ai = 1, and up to considering a linear combina-
tion of y1 and y2 , we can assume that φ1 = 0 and φ2 = − π2 . Thus, we have w0 = 1. This
concludes the proof.

167
 Appendix B

S OME TECHNICAL RESULTS

We collect here some technical results used in our proofs. Recall first a sufficient
condition for the non-explosion of the solution to a SDE. The proof can be found in
[GL21a].

Lemma B.0.1. Let (Yt )t≥t0 be a càdlàg process, solution to a SDE. For all n ≥ 0, define
the stopping time
τn := inf{t ≥ t0 , kYt k ≥ n}. (B.1)

Set τ∞ := limn→+∞ τn the explosion time of Y . Assume that there exist two measurable
and non-negative functions φ and b such that
(i) φ is non-decreasing and limn→∞ φ(n) = +∞,
(ii) b is finite-valued,
(iii) and for all t ≥ t0 ,
sup E [φ(|Yt∧τn |)] ≤ b(t).
n≥0

Then τ∞ = +∞ a.s.

We now state and prove a result on the periodic-averaging phenomenon.

Lemma B.0.2. Let us fix t0 > 0 and h : [t0 , +∞) → R a continuous m-periodic function,
with m > 0. Let g : [t0 , +∞) → R+ be a continuously differentiable function which is not
integrable on [t0 , +∞). We assume moreover that
(i) g 0 (t) = o (g(t)),
t→+∞
R 
t
(ii) g(t) = o t0 g(u) du .
t→+∞
Then,
Z t
1 Z t0 +m
 Z t Z t 
g(u)h(u) du = h(u) du g(u) du + o g(u) du .
t0 m t0 t0 t→+∞ t0

168
Let us remark that the functions g1 and g2 defined for t ∈ R, r ≥ 0 and β ∈ [0, 1) by
g1 (t) := tr and g2 (t) := exp(rt1−β ), satisfy the preceding assumptions made on g.
e := h − 1 R t0 +m h(u) du, and H
Proof. Let us define h f a primitive of h.
e The function Hf
m t0
is bounded on [t0 , +∞) since the average of he on its period is equal to 0. To prove the

lemma, we only need to justify that

Z t Z t 
g(u)h(u)
e du = o g(u) du .
t0 t→+∞ t0

By integration by parts, we obtain that, for all t ≥ t0 ,

Z t Z t
g(u)h(u)
e du = g(t)H(t)
f − g(t0 )H(t
f )−
0 g 0 (u)H(u)
f du.
t0 t0

f is bounded, that g 0 (t) = o (g(t)) and that R ∞ g(u) du = +∞,

Using the fact that H
t→+∞ t0
we deduce that Z t Z t 
0
g (u)H(u) du = o
f g(u) du .
t0 t→+∞ t0
R 
t
The conclusion follows from the fact that g(t)H(t)
f − g(t0 )H(t
f ) =
0 o t0 g(u) du ,
R  t→+∞
t R∞
since we have assumed that g(t) = o t0 g(u) du and that t0 g(u) du = +∞.
t→+∞

 i
1
Lemma B.0.3. Let f be given by (A.2) for β ∈ 2
,1 , and pick α ∈ (0, 2]. Define

(1 + α/2)−1

if β = 1,
kβ,α := 
2 else.
α

Then for any t > 0, we have

Z t/ε
f (u)−α u1−2β du = O (f (t/ε)−α εβ−1 ),
t0 ε→0

and Z t/ε  −α  β

−α t t
f (u) du = kβ,α f + o (f (t/ε)−α ε−β ).
t0 ε ε ε→0

Proof. When β = 1, the results follow from direct computations because of the expression
 
of f . Assume now that β ∈ 21 , 1 . For the first point, the integration by parts formula

169
Chapter B – Some technical results

ensures that
Z t/ε
2h it/ε 2 Z t/ε
f (u)−α u1−2β du = f (u)−α u1−β − (1 − β) f (u)−α u−β du
t0 α t0 α t0

= O (f (t/ε)−α εβ−1 ) + O (f (t/ε)−α )

ε→0 ε→0
−α β−1
= O (f (t/ε) ε ).
ε→0

For the second asymptotic expansion, it follows again from an integration by parts that
Z t/ε Z t/ε
−α
f (u) du = f (u)−α u−β uβ du
t0 t0
2h it/ε 2 Z t/ε
= f (u)−α uβ − β f (u)−α uβ−1 du.
α t0 α t0

Remarking that f (u)−α uβ−1 = o (f (u)−α ), since β < 1, we deduce that

u→+∞

Z t/ε Z t/ε !
−α β−1 −α
f (u) u du = o f (u) du .
t0 ε→0 t0

We obtain that Z t/ε  −α  β

−α 2 t t
f (u) du ∼ f
t0 ε→0 α ε ε
This ends the proof.
Lemma B.0.4. Let (Xn )n be a sequence of random variables with values in R2 , and
which converges in distribution to a random variable X. We assume that the distribution
of X is invariant under rotations, i.e. for any orthogonal matrix R ∈ M2 (R), the random
variables X and RX have the same distribution. Then for all sequence (Rn )n of orthogonal
matrices in M2 (R), we have
Rn Xn =⇒ X.
n→+∞
Proof. Let us denote by φZ the characteristic function of a random variable Z. Using
Theorem 5.3 p. 86 in [Kal02], we know that (φXn )n converges to φX uniformly on every
compact subset of R2 . The characteristic function of the random variable Yn := Rn Xn is
given by
ξ 7→ φYn (ξ) = φXn (Rn ξ).

Thus, by assumption, we have, for all ξ ∈ R2 ,

φX (Rn ξ) = φX (ξ).

170
It follows that, for any ξ ∈ R2 and n ≥ 0,

|φYn (ξ) − φ(ξ)| = |φXn (Rn ξ) − φX (Rn ξ)| ≤ sup |φXn (z) − φX (z)| ,
z∈R2 ,kzk=kξk

which converges to 0, as n → +∞. This ends the proof of the lemma.

Acknowledgements
The authors would like to thank Paul-Eric Chaudru de Raynal and Mihai Gradinaru
for their supervision and advice. The authors would also like to thank Jürgen Angst for
having suggested this subject and useful discussions.

171
 B IBLIOGRAPHY

[AK94] S. Albeverio and A. Klar, « Long Time Behavior of Nonlinear Stochastic Os-
cillators: The One-dimensional Hamiltonian Case », in: J. Math. Phys. 35.8
(1994), pp. 4005–4027, doi: 10.1063/1.530839 (cit. on pp. 34, 128).
[App09] D. Applebaum, Levy Processes and Stochastic Calculus, Second, Cambridge
Studies in Advanced Mathematics, Cambridge University Press, 2009 (cit. on
pp. 24, 32, 91–93, 138, 149, 150, 152, 153).
[AS09] D. Applebaum and M. Siakalli, « Asymptotic Stability of Stochastic Differen-
tial Equations Driven by Lévy Noise », in: Journal of Applied Probability 46.4
(2009), pp. 1116–1129, doi: 10.1239/jap/1261670692 (cit. on p. 80).
[AW09] J. Appleby and H. Wu, « Solutions of Stochastic Differential Equations Obey-
ing the Law of the Iterated Logarithm, with Applications to Financial Mar-
kets », in: Electron. J. Probab. 14 (2009), pp. 912–959, doi: 10.1214/EJP.v14-
642 (cit. on pp. 12, 81).
[BGL14] D. Bakry, I. Gentil, and M. Ledoux, Analysis and Geometry of Markov Dif-
fusion Operators, vol. 348, Grundlehren Der Mathematischen Wissenschaften,
Cham: Springer International Publishing, 2014, doi: 10.1007/978- 3- 319-
00227-9 (cit. on pp. 30, 69).
[Bas03] R. F. Bass, « Stochastic Differential Equations Driven by Symmetric Stable
Processes », in: Séminaire de Probabilités XXXVI, Lecture Notes in Mathe-
matics, Springer, 2003, pp. 302–313, doi: 10.1007/978-3-540-36107-7_11
(cit. on pp. 8, 80).
[Ber21] N. Berglund, Long-Time Dynamics of Stochastic Differential Equations, 2021,
arXiv: 2106.12998 (cit. on p. 6).
[Bét21] L. Béthencourt, Stable Limit Theorems for Additive Functionals of One Di-
mensional Diffusion Processes, 2021, arXiv: 2104.06027 (cit. on pp. 15, 46).
[BY87] P. Biane and M. Yor, « Valeurs Principales Associées Aux Temps Locaux
Browniens », in: Bull. Sci. Math. (2) 111.1 (1987), pp. 23–101 (cit. on p. 15).

173
[Bil99] P. Billingsley, Convergence of Probability Measures, 2nd ed, Wiley Series in
Probability and Statistics. Probability and Statistics Section, Wiley, 1999 (cit.
on pp. 17, 29, 39, 60, 63, 67, 70, 94, 95, 100–102, 118, 138, 140, 141, 153, 155,
157).
[Bob99] S. G. Bobkov, « Isoperimetric and Analytic Inequalities for Log-Concave Prob-
ability Measures », in: The Annals of Probability 27.4 (1999), pp. 1903–1921
(cit. on pp. 30, 69).
[Bre68] L. Breiman, « A Delicate Law of the Iterated Logarithm for Non-Decreasing
Stable Processes », in: Ann. Math. Statist. 39.6 (1968), pp. 1818–1824, doi:
10.1214/aoms/1177698014 (cit. on p. 12).
[CCM10] P. Cattiaux, D. Chafai, and S. Motsch, « Asymptotic Analysis and Diffu-
sion Limit of the Persistent Turning Walker Model », in: Asymptotic Analysis
67.1–2 (2010), pp. 17–31, doi: 10.3233/ASY-2009-0969, arXiv: 0811.0600
(cit. on pp. 5, 6, 16, 29, 30, 46, 69).
[CNP19] P. Cattiaux, E. Nasreddine, and M. Puel, « Diffusion Limit for Kinetic Fokker-
Planck Equation with Heavy Tails Equilibria: The Critical Case », in: Kinetic
& Related Models 12.4 (2019), p. 727, doi: 10.3934/krm.2019028 (cit. on
pp. 14, 46).
[CL] T. Cavallazzi and E. Luirard, Asymptotic Behavior of a Time-Inhomogeneous
Stochastic System in the Quadratic Potential (cit. on pp. 2, 47).
[CM20] P.-E. Chaudru De Raynal and S. Menozzi, On Multidimensional Stable-Driven
Stochastic Differential Equations with Besov Drift, 2020, arXiv: 1907.12263
(cit. on pp. 24, 81, 90).
[CG00] A. V. Chechkin and V. Y. Gonchar, « Linear Relaxation Processes Governed
by Fractional Symmetric Kinetic Equations », in: J. Exp. Theor. Phys. 91.3
(2000), pp. 635–651, doi: 10.1134/1.1320100, arXiv: cond- mat/9910091
(cit. on p. 7).
[CSZ18] Z.-Q. Chen, R. Song, and X. Zhang, « Stochastic Flows for Lévy Processes
with Hölder Drifts », in: Revista Matemática Iberoamericana 34 (2018), doi:
10.4171/rmi/1042, arXiv: 1501.04758 (cit. on p. 80).

174
[CZZ17] Z.-Q. Chen, X. Zhang, and G. Zhao, Well-Posedness of Supercritical SDE
Driven by Lévy Processes with Irregular Drifts, 2017, arXiv: 1709.04632 (cit.
on pp. 21, 80, 86).
[CE05] A. S. Cherny and H. J. Engelbert, Singular Stochastic Differential Equations,
Lecture Notes in Mathematics 1858, Springer, 2005 (cit. on pp. 52, 63).
[DF13] A. Debussche and N. Fournier, « Existence of Densities for Stable-like Driven
SDE’s with Hölder Continuous Coefficients. », in: Journal of Functional Anal-
ysis 264 (2013), pp. 1757–1778, doi: 10.1016/[Link].2013.01.009 (cit. on
p. 154).
[DM08] P. Degond and S. Motsch, « Large Scale Dynamics of the Persistent Turning
Walker Model of Fish Behavior », in: J Stat Phys 131.6 (2008), pp. 989–1021,
doi: 10.1007/s10955-008-9529-8 (cit. on p. 5).
[DS15] C.-S. Deng and R. L. Schilling, « On Shift Harnack Inequalities for Subordi-
nate Semigroups and Moment Estimates for Lévy Processes », in: Stochastic
Processes and their Applications 125.10 (2015), pp. 3851–3878, doi: 10.1016/
[Link].2015.05.013 (cit. on pp. 24, 81, 91, 111, 112).
[Dit99] P. D. Ditlevsen, « Observation of α-Stable Noise Induced Millennial Climate
Changes from an Ice-Core Record », in: Geophysical Research Letters 26 (1999),
pp. 1441–1444, doi: 10.1029/1999GL900252 (cit. on pp. 7, 80, 81).
[Don18] Y. Dong, « Jump Stochastic Differential Equations with Non-Lipschitz and
Superlinearly Growing Coefficients », in: Stochastics 90.5 (2018), pp. 782–
806, doi: 10.1080/17442508.2017.1421195 (cit. on p. 80).
[EG15] R. Eon and M. Gradinaru, « Gaussian Asymptotics for a Non-Linear Langevin
Type Equation Driven by an α-Stable Lévy Noise », in: Electron. J. Probab.
20 (2015), doi: 10.1214/EJP.v20-4068 (cit. on pp. 16, 17, 46).
[Fed+17] E. Fedrizzi et al., « Regularity of Stochastic Kinetic Equations », in: Electron.
J. Probab. 22 (2017), doi: 10.1214/17-EJP65 (cit. on pp. 34, 129, 131).
[FT21] N. Fournier and C. Tardif, « One Dimensional Critical Kinetic Fokker-Planck
Equations, Bessel and Stable Processes », in: Communications in Mathemat-
ical Physics 381 (2021), doi: 10.1007/s00220-020-03903-0, arXiv: 1805.
09728 (cit. on pp. 14–16, 46, 62, 81, 84, 128).

175
[GL21a] M. Gradinaru and E. Luirard, Asymptotic Behavior for a Time-Inhomogeneous
Kolmogorov Type Diffusion, 2021, arXiv: 2004.11576 (cit. on pp. 2, 45, 80,
82, 103, 118, 128–131, 135, 136, 148, 168).
[GL21b] M. Gradinaru and E. Luirard, Asymptotic Behavior for a Time-Inhomogeneous
Stochastic Differential Equation Driven by an α-Stable Lévy Process, 2021,
arXiv: 2112.07287 (cit. on pp. 2, 47, 79, 128–131, 148, 149, 154).
[GO13] M. Gradinaru and Y. Offret, « Existence and Asymptotic Behaviour of Some
Time-Inhomogeneous Diffusions », in: Ann. Inst. H. Poincaré Probab. Statist.
49.1 (2013), pp. 182–207, doi: 10.1214/11-AIHP469 (cit. on pp. 12, 13, 21,
28, 30, 33, 46–50, 52–54, 56, 64, 71, 74, 75, 81, 84, 96, 128, 129).
[HMC18] I. Honore, S. Menozzi, and P.-E. Chaudru de Raynal, Strong Regularization
by Brownian Noise Propagating through a Weak Hörmander Structure, 2018,
arXiv: 1810.12225 [math] (cit. on pp. 34, 38, 129, 135).
[IW81] N. Ikeda and S. Watanabe, Stochastic Differential Equations and Diffusion
Processes, North-Holland Mathematical Library 24, Elsevier, Academic Press,
1981 (cit. on pp. 21, 86, 148).
[JS03] J. Jacod and A. N. Shiryaev, The General Theory of Stochastic Processes,
Semimartingales and Stochastic Integrals, Grundlehren Der Mathematischen
Wissenschaften, Berlin, Heidelberg: Springer, 2003, doi: 10 . 1007 / 978 - 3 -
662-05265-5_1 (cit. on pp. 11, 16–18, 107).
[Kal02] O. Kallenberg, Foundations of Modern Probability, Probability and Its Appli-
cations, New York, NY: Springer New York, 2002, doi: 10 . 1007 / 978 - 1 -
4757-4015-8 (cit. on pp. 13, 29, 39, 64, 65, 74, 135, 139, 143, 170).
[KS98] I. Karatzas and S. Shreve, Brownian Motion and Stochastic Calculus, Second,
Graduate Texts in Mathematics, New York: Springer-Verlag, 1998, doi: 10.
1007/978-1-4612-0949-2 (cit. on pp. 13, 52, 63, 66, 120, 121, 137).
[Kol29] A. Kolmogoroff, « Über das Gesetz des iterierten Logarithmus », in: Mathe-
matische Annalen 101 (1929), pp. 126–135 (cit. on p. 10).
[Küh] F. Kühn, Lévy Matters VI - Lévy-Type Processes: Moments, Construction and
Heat Kernel Estimates, Springer (cit. on p. 41).

176
[Küh17] F. Kühn, « Existence and Estimates of Moments for Lévy-type Processes »,
in: Stochastic Processes and their Applications 127.3 (2017), pp. 1018–1041,
doi: 10.1016/[Link].2016.07.008 (cit. on p. 81).
[Kul09] A. M. Kulik, « Exponential Ergodicity of the Solutions to SDE’s with a Jump
Noise », in: Stochastic Processes and their Applications 119.2 (2009), pp. 602–
632, doi: 10.1016/[Link].2008.02.006 (cit. on p. 98).
[Kun04] H. Kunita, « Stochastic Differential Equations Based on Lévy Processes and
Stochastic Flows of Diffeomorphisms », in: Real and Stochastic Analysis: New
Perspectives, Trends in Mathematics, Boston, MA: Birkhäuser, 2004, pp. 305–
373, doi: 10.1007/978-1-4612-2054-1_6 (cit. on p. 116).
[Kur07] V. P. Kurenok, « Stochastic Equations with Time-Dependent Drift Driven by
Levy Processes », in: J Theor Probab 20.4 (2007), pp. 859–869, doi: 10.1007/
s10959-007-0086-x (cit. on pp. 80, 117).
[Lan08] P. Langevin, « Sur la théorie du mouvement brownien », in: Comptes-Rendus
de l’Académie des Sciences 146 (1908), pp. 530–532 (cit. on p. 4).
[LP19] G. Lebeau and M. Puel, « Diffusion Approximation for Fokker Planck with
Heavy Tail Equilibria: A Spectral Method in Dimension 1 », in: Commun.
Math. Phys. 366.2 (2019), pp. 709–735, doi: 10.1007/s00220-019-03315-9
(cit. on p. 14).
[Lui22a] E. Luirard, Simulation for "Asymptotic Behavior for a Time-Inhomogeneous
Kolmogorov Type Diffusion", 2022, doi: 10.5281/zenodo.6472854 (cit. on
pp. 21, 22, 27).
[Lui22b] E. Luirard, Simulation for "Asymptotic Behavior for a Time-Inhomogeneous
Stochastic Differential Equation Driven by an α-Stable Lévy Process", 2022,
doi: 10.5281/zenodo.6472878 (cit. on pp. 21, 23).
[Lui22c] E. Luirard, Simulation for "Asymptotic Behavior of a Time-Inhomogeneous
Stochastic System in the Quadratic Potential ", 2022, doi: 10.5281/zenodo.
6472888 (cit. on pp. 36, 37).
[LP06] H. Luschgy and G. Pagès, « Moment Estimates for Lévy Processes », in: Elec-
tronic Communications in Probability 13 (2006), doi: 10 . 1214 / ECP . v13 -
1397, arXiv: math/0607282 (cit. on pp. 24, 81).

177
[MM21] L. Marino and S. Menozzi, Weak Well-Posedness for Degenerate SDEs Driven
by Lévy Processes, 2021, arXiv: 2107.04325 (cit. on pp. 34, 38, 129, 148, 149).
[NP15] E. Nasreddine and M. Puel, « Diffusion Limit of Fokker-Planck Equation with
Heavy Tail Equilibria », in: ESAIM: M2AN 49.1 (2015), pp. 1–17, doi: 10.
1051/m2an/2014020 (cit. on p. 14).
[Off14] Y. Offret, « Invariant Distributions and Scaling Limits for Some Diffusions
in Time-Varying Random Environments », in: Probab. Theory Relat. Fields
158.1 (2014), pp. 1–38, doi: 10.1007/s00440-012-0475-7 (cit. on pp. 10,
21, 33, 42).
[Pil13] A. Pilipenko, « On Existence and Properties of Strong Solutions of One-
Dimensional Stochastic Equations with an Additive Noise », in: Theory of
Stochastic Processes 18.2 (2013), arXiv: 1306.0212 (cit. on p. 80).
[Pri+12] E. Priola et al., « Exponential Ergodicity and Regularity for Equations with
Lévy Noise », in: Stochastic Processes and their Applications 122.1 (2012),
pp. 106–133, doi: 10.1016/[Link].2011.10.003 (cit. on p. 80).
[Pro05] P. E. Protter, Stochastic Integration and Differential Equations, Second, Sto-
chastic Modelling and Applied Probability, Berlin Heidelberg: Springer-Verlag,
2005, doi: 10.1007/978-3-662-10061-5 (cit. on pp. 88, 91, 117, 149, 152).
[Rek20] J. Reker, Short-Time Behavior of Solutions to Lévy-driven SDEs, 2020, arXiv:
2008.00526 (cit. on p. 80).
[RY05] D. Revuz and M. Yor, Continuous Martingales and Brownian Motion, Berlin;
Heidelberg: Springer, 2005 (cit. on pp. 13, 53).
[RT11] M. Rosenbaum and P. Tankov, « Asymptotic Results for Time-Changed Lévy
Processes Sampled at Hitting Times », in: Stochastic Processes and their Ap-
plications 121.7 (2011), pp. 1607–1632, doi: 10.1016/[Link].2011.03.013
(cit. on pp. 31, 106).
[SG03] G. Samorodnitsky and M. Grigoriu, « Tails of Solutions of Certain Nonlinear
Stochastic Differential Equations Driven by Heavy Tailed Lévy Motions »,
in: Stochastic Processes and their Applications 105.1 (2003), pp. 69–97, doi:
10.1016/S0304-4149(03)00002-4 (cit. on p. 135).
[ST94] G. Samorodnitsky and M. S. Taqqu, Stable Non-Gaussian Random Processes:
Stochastic Models with Infinite Variance, CRC Press, 1994 (cit. on pp. 28, 96).

178
[Sat99] K.-i. Sato, Lévy Processes and Infinitely Divisible Distributions, Cambridge
Studies in Advanced Mathematics 68, Cambridge, U.K. ; New York: Cam-
bridge University Press, 1999 (cit. on pp. 104, 107, 145, 156, 159).
[Sit05] R. Situ, Theory of Stochastic Differential Equations with Jumps and Applica-
tions: Mathematical and Analytical Techniques with Applications to Engineer-
ing, Mathematical and Analytical Techniques with Applications to Engineer-
ing, Springer US, 2005, doi: 10.1007/b106901 (cit. on pp. 21, 86, 148).
[SV06] D. W. Stroock and S. R. S. Varadhan, Multidimensional Diffusion Processes,
Classics in Mathematics, Berlin Heidelberg: Springer-Verlag, 2006, doi: 10.
1007/3-540-28999-2 (cit. on pp. 53–56).
[WZ15] F.-Y. Wang and X. Zhang, Degenerate SDEs in Hilbert Spaces with Rough
Drifts, 2015, arXiv: 1501.04150 (cit. on pp. 34, 129).
[Wat44] G. N. Watson, A Treatise on the Theory of Bessel Functions, Second, Cambrige
University Press, 1944 (cit. on p. 167).
[Wu01] L. Wu, « Large and Moderate Deviations and Exponential Convergence for
Stochastic Damping Hamiltonian Systems », in: Stochastic Processes and their
Applications 91.2 (2001), pp. 205–238, doi: 10.1016/S0304-4149(00)00061-
2 (cit. on pp. 34, 128).
[Yam73] T. Yamada, « On a Comparison Theorem for Solutions of Stochastic Differ-
ential Equations and Its Applications », in: Journal of Mathematics of Kyoto
University 13.3 (1973), pp. 497–512, doi: 10.1215/kjm/1250523321 (cit. on
pp. 74, 75).
[Yor08] M. Yor, Du pli cacheté de Döblin aux équations différentielles stochastiques :
conférence du mercredi 23 janvier 2008, 2008 (cit. on p. 4).
[Zha13] X. Zhang, « Stochastic Differential Equations with Sobolev Drifts and Driven
by α-Stable Processes », in: Ann. Inst. H. Poincaré Probab. Statist. 49.4 (2013),
pp. 1057–1079, doi: 10.1214/12-AIHP476 (cit. on p. 80).
[Zha14] X. Zhang, « Densities for SDEs Driven by Degenerate α-Stable Processes »,
in: The Annals of Probability 42.5 (2014), pp. 1885–1910, doi: 10.1214/13-
AOP900 (cit. on pp. 34, 129).
[Zha16] X. Zhang, Stochastic Hamiltonian Flows with Singular Coefficients, 2016, arXiv:
1606.04360 (cit. on pp. 34, 129).

179
Titre : Loi limite de modèles cinétiques inhomogènes en temps

Mot clés : modèle stochastique cinétique ; équation différentielle stochastique inhomogène en

temps ; processus de Lévy ; loi limite ; transformation d’échelles

Résumé : Nous étudions, dans cette thèse, La première partie de ce manuscrit corres-
le comportement asymptotique de solutions pond à l’étude du système en l’absence de
de systèmes cinétiques inhomogènes, dirigés potentiel confinant, tandis que la seconde s’in-
par un processus de Lévy L. Plus précisé- téresse à la présence d’un potentiel quadra-
ment, on s’intéresse à la dynamique d’une tique.
particule, évoluant dans un potentiel U, et sou- L’enjeu est de comprendre comment inter-
mise à la fois à une force de frottement F et à agissent les différentes forces afin de mon-
une force extérieure aléatoire L. La force F est trer que le processus vitesse-position, correc-
attractive et vérifie des propriétés d’invariance tement renormalisé, admet une limite en loi
d’échelle. Elle est altérée par la présence d’un explicite.
facteur t−β .

Title: Asymptotic distribution of a time-inhomogeneous kinetic model

Keywords: kinetic stochastic equation; time-inhomogeneous stochastic differential equation;

Lévy process; asymptotic distribution; scaling transformation

Abstract: We study the asymptotic behav- tem. Alternatively, in the second part, the par-
ior of some stochastic kinetic inhomogeneous ticle is supposed to evolve in a quadratic po-
models driven by a Lévy process L. It de- tential.
scribes the dynamics of a particle, evolving in The issue of this thesis is to understand the
a potential U, subject to a random force L and balance between the drag force and the ran-
a frictional force F . The drag force is supposed dom force, with the aim of highlighting an ex-
to satisfy some scaling properties and to be plicit distributional limit of the couple velocity-
weakened along the time. position.
In the first part, we study the free potential sys-