Étude de Modèles Mathématiques Des Condensats de Bose-Einstein Pour Différents Types de Pièges Et D'interactions - Jimena Royo-Letelier

Étude de modèles mathématiques des
condensats de Bose-Einstein pour

différents types de pièges et d’interactions
THÈSE DE DOCTORAT DE L’UNIVERSITÉ DE
VERSAILLES-SAINT-QUENTIN-EN-YVELINES
Présentée par :
Jimena ROYO-LETELIER
Pour obtenir le grade de docteur de l’Université de Versailles
Soutenue le 10 juin 2013, devant le jury composé de :
Amandine AFTALION Directrice de thèse

Dorin BUCUR Rapporteur
María J. ESTEBAN Examinatrice
Bernard HELFFER Directeur de thèse
Etienne SANDIER Rapporteur
Susanna TERRACINI Examinatrice
2
3
Résumé
Cette thèse porte sur l’étude mathématique de modèles théoriques des condensats
de Bose-Einstein. On considère la fonctionnelle d’énergie de Gross-Pitaevskii pour
différents types de piégeages et d’interactions. On étudie des modèles de condensats
à deux dimensions définis sur tout l’espace, en rotation et à plusieurs composants,
ainsi qu’un modèle décrivant une particule chargée dans un milieu périodique bi-
dimensionnel avec champ magnétique. Les outils mathématiques utilisés sont les
équations aux dérivées partielles, l’analyse non linéaire, la théorie géométrique de la
mesure, la théorie spectrale et l’analyse semi-classique. Les résultats principaux vont
dans quatre directions. Le premier résultat établit la non existence de vortex dans la
zone de faible densité d’un condensat en rotation sous-critique. Le deuxième résultat
montre la brisure de symétrie et de la ségrégation d’un condensat à deux composants
dans le régime de fort couplage et faible interaction. On résout aussi un problème
de partition optimale spectrale associée à un opérateur de Schrödinger dans le plan.
On introduit un nouveau modèle de minimisation du périmètre pour l’étude d’un
condensat à deux composants dans le régime de fort couplage et forte interaction.
Le troisième résultat concerne la Γ-convergence de la fonctionnelle d’énergie d’un
condensat à deux composants dans ce dernier régime. On montre aussi la brisure
de symétrie d’un condensat à deux composants dans le régime de fort couplage et
forte interaction. Le dernier résultat traite du spectre d’un opérateur de Schrödinger
périodique magnétique dans un réseau de kagome.
Abstract
This PhD thesis is devoted to the mathematical study of theoretical models for
Bose-Einstein condensates. We consider the Gross-Pitaevskii functional for several
types of trapping potentials and interactions. We analyze models for two-dimensional
condensates defined over all R2 , under rotation and with several components. We
also analyze a model for a charged particle in a two-dimensional periodic me-
dia under magnetic field. The mathematical tools employed are partial differen-
tial equations, nonlinear analysis, geometric measure theory, spectral theory and
semi-classical analysis. The are four main results. The first one establishes the non
existence of vortex in the low density zone of a condensate under subcritical ro-
tation. The second result proves the segregation and the symmetry breaking of a
two-component condensate in the strongly coupled and weakly interacting regime.
We also solve an optimal partition problem associated with a Schrödinger operator
in R2 . We introduce a new minimal perimeter model for the study of two-component
condensate in the strongly coupled and strongly interacting regime. The third result
is about the Γ-convergence of the energy functional of a two-component condensate
in this last regime. We also show the symmetry breaking of a two-component conden-
sate in the strongly coupled and strongly interacting regime. The last result concerns
the spectrum of a magnetic periodical Schrödinger operator on the kagome lattice.
4
5
A Emilie
A ma famille
¡M«as vale ser punky !

6
7
ENIVREZ-VOUS
Il faut être toujours ivre.

Tout est là :
c’est l’unique question.
Pour ne pas sentir
l’horrible fardeau du Temps
qui brise vos épaules et vous penche vers la terre,
il faut vous enivrer sans trêve.
Mais de quoi ?
De vin, de poésie ou de vertu 1 , à votre guise.
Mais enivrez-vous.
Et si quelquefois,
sur les marches d’un palais,
sur l’herbe verte d’un fossé,
dans la solitude morne de votre chambre,
vous vous réveillez,
l’ivresse déjà diminuée ou disparue,
demandez au vent,
à la vague,
à l’étoile,
à l’oiseau,
à l’horloge,
à tout ce qui fuit,
à tout ce qui gémit,
à tout ce qui roule,
à tout ce qui chante,
à tout ce qui parle,
demandez quelle heure il est,
et le vent,
la vague,
l’étoile,
l’oiseau,
l’horloge,
vous répondront :
« Il est l’heure de s’enivrer !
Pour n’être pas les esclaves martyrisés du Temps,
enivrez-vous ;
enivrez-vous sans cesse !
De vin, de poésie ou de vertu, à votre guise. »
Charles Baudelaire
Le Spleen de Paris, XXXIII
1. ou de mathématiques !
8
9
Remerciements
Je voudrais remercier mes directeurs, Amandine Aftalion et Bernard Helffer, d’avoir

dirigé ma thèse et guidé mes premiers pas dans le monde de la recherche. Je leur suis
reconnaissante par les très beaux et passionnants sujets de recherche qu’ils m’ont
proposés. Travailler avec eux a toujours été enrichissant pour moi et je tiens à les
remercier tous les deux pour leur très grande disponibilité. Je remercie Amandine
pour les nombreuses opportunités qu’elle m’a permis d’avoir et en particulier pour
m’avoir fait connaître la physique des condensats de Bose-Einstein. Je tiens à lui
exprimer ma gratitude pour sa confiance et ses encouragements. Je remercie Bernard
pour son infinie patience et sa cordialité et pour avoir partagé avec moi son immense
savoir mathématique. Je tiens à le remercier pour m’avoir encouragée pendant la fin
de ma thèse et pour m’avoir accueillie chez lui à Nantes à plusieurs reprises pendant
cette période.
Je remercie très sincèrement Dorin Bucur et Etienne Sandier de m’avoir fait l’hon-
neur de rapporter cette thèse. C’est un grand plaisir pour moi de les avoir dans
mon jury. Je suis également très honorée que María Esteban et Susanna Terracini
aient accepté de faire partie de mon jury. Je remercie chaleureusement Susanna pour
l’intérêt qu’elle a porté à mes travaux et pour les nombreuses discussions que nous
avons partagées.
Au cours de ma thèse j’ai eu l’opportunité de visiter Robert L. Jerrard à l’Univer-

sity of Toronto. Je le remercie pour cette invitation et pour son hospitalité ainsi
que pour l’intérêt qu’il a porté à mes travaux et pour son soutien. J’ai beaucoup
profité de cette expérience et des discussions mathématiques avec Robert. Aussi, j’ai
eu la chance d’être accueillie en visite au Departamento de Física de la Pontificia
Universidad Católica de Chile. Je remercie Rafael Benguria pour cette invitation et
son accueil chaleureux, ainsi que pour son intérêt pour mon travail et son soutien.
Ça a été très enrichissant de discuter avec les physiciens du Departamento. Merci
également à Robert Seiringer pour son invitation au Department of Mathematics
and Statistics de McGill University.
Je suis reconnaissante aussi envers les personnes qui ont partagé leur savoir ma-
thématique et/ou physique avec moi de façon plus au moins informelle. La beauté
des théorèmes mathématiques prend tout son sens lorsqu’on s’aperçoit qu’ils sont
le fruit des échanges et discussions d’une communauté de gens qui aiment parta-
ger leurs connaissances de façon désintéressée. Je tiens à remercier tous les amis et
collègues avec qui j’ai pu échanger des idées mathématiques ces dernières années,
spécialement à Chico, Coni, Michael et Peter. Je remercie aussi Guy Bouchitté, Juan
Dávila, Manuel del Pino, Clément Gallo, Duván Henao, Loïc Le Treust, Peter Ma-
son, Benedetta Noris, Philippe Kerdelhue, Dominique Spehner et Juncheng Wei.
10
Pendant la dernière année j’ai été lauréate de la bourse Bourses L’Oréal-France -

UNESCO « Pour les Femmes et la Science ». Je remercie ces institutions pour cette
récompense, qui m’a permis de finir ma thèse dans des conditions confortables, ainsi
que d’organiser une rencontre entre jeunes chercheurs en mathématiques et en phy-
sique. Je remercie aussi tous les orateurs et à l’Institut Henri Poincaré pour avoir
accueilli cette rencontre.
Ce mémoire marque la fin de ma thèse de doctorat, mais il représente aussi l’abou-

tissement des dix années d’études. Je me permets donc de profiter de cet espace
pour remercier tous ceux qui ont participé à mon éducation. Tout premièrement à
ma famille, à qui je dédie des mots un peu plus loin. Je remercie vivement les profs
de nombreux lycées que j’ai fréquentés qui ont su voir en moi un peu plus qu’une
adolescente révoltée et qui ont cru en moi. J’ai une pensée toute particulière pour
Hugo Cepeda, j’espère pouvoir lui faire parvenir ce mémoire de thèse où qu’il soit.
Je remercie aussi mes profs à Las Palmeras de la Facultad des Ciencias de l’Uni-
versidad de Chile. Ils m’ont montré pour la première fois l’harmonie et l’élégance
des mathématiques et de la physique et m’ont encouragée à suivre la voie de la
recherche. Je remercie notamment Andrés Navas pour son amitié et son soutien. Je
pense à mes camarades de Las Palmeras avec qui j’ai eu le plaisir de partager des
discussions mêlant sciences, philosophie et art, le tout dans les agréables et vertes
pelouses du Campus Juan Gomez Millas. Merci à Gabriel et Ernesto avec qui je
continue ces motivantes conversations en France. Je suis heureuse d’avoir commencé
mes études à l’Universidad de Chile, pour laquelle j’ai une affection particulière de-
puis mes premiers pas (littéralement !) à Antumapu. J’espère avec tout mon coeur
que la suite des événements au Chili saura restaurer ses principes fondamentaux et
que tous ceux qui le méritent pourront y faire ses études gratuitement comme mon
père l’a fait un jour. Toujours dans l’Amérique du Sud, je garde un très bon souve-
nir de mon bref passage d’été à l’Instituto Nacional de Matemática Pura e Aplicada
à Rio de Janeiro. Finalement, je tiens à remercier d’une façon un peu abstraite le
système éducatif français. J’ai eu la chance de réaliser huit ans de mes études en
France, toujours dans un cadre très confortable qui m’a permis de travailler dans
des conditions optimales. Je suis très reconnaissante et j’espère avoir honoré cette
précieuse opportunité qui m’a été donnée.
Je remercie également Ayrin, David et Mayte pour leur compagnie, leur soutien et
leur bonne humeur à l’École Polytechnique. À mes parrains de l’X, Rouquès, Charles
et Paula, pour leur accueil en France, qui a été très important pour moi. Je remercie
aussi Julien pour son soutien au début de ma thèse. Je suis infiniment reconnais-
sante envers mes camarades de vie : Florencia, Jorge et Ignacio pour m’avoir appris
tant de choses. Malgré la distance je suis heureuse de vous garder toujours dans
mon cœur. Je remercie aussi mes camarades salseros, ce fut un grand plaisir d’avoir
partagé trois ans de musique avec vous !
11
Je laisse pour la fin les plus importants des remerciements.
Les mots me manquent pour exprimer ma reconnaissance envers Émilie. Je ne pour-

rais pas la remercier assez pour tout son amour, compréhension et encouragement,
surtout dans les moments difficiles de cette thèse. Ce mémoire lui est dédié avec tout
mon amour.
Enfin, je voudrais remercier ma famille, dans les mots de ma soeur Manuela, un

exemple d’amour et liberté. Je tiens à exprimer gratitude à mon père et ma mère
pour leurs constants soutien et amour, pour m’avoir appris que tout était possible.
À mon père pour m’avoir transmis le goût des mathématiques et de la nature et
pour m’avoir appris l’importance de la solidarité. A ma mère pour m’avoir transmis
le goût des arts et m’avoir appris l’importance du respect des gens. À ma sœur,
qui est un exemple à suivre pour moi, pour son encouragement et pour tous ses
enseignements. Si je suis aujourd’hui en train de défendre ma thèse de doctorat,
c’est grâce à l’aide de ma soeur qui a pris en charge mes responsabilités au Chili.
Je lui suis infiniment reconnaissante. À Andrés pour tout son soutien et son amour.
Je n’aurais jamais pu faire une thèse en France sans leurs support et encouragement.
Jimena
12
Table des matières
1 Introduction 17
1.1 La physique des condensats de Bose-Einstein . . . . . . . . . . . . . . 17
1.1.1 La condensation de Bose-Einstein . . . . . . . . . . . . . . . . 17
1.1.2 L’approximation de Hartree et l’énergie de
Gross-Pitaevskii . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.1.3 Configurations expérimentales . . . . . . . . . . . . . . . . . . 20
1.1.4 Magnétisme artificiel avec des condensats de
Bose-Einstein . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.2 La fonctionnelle de Gross-Pitaevskii . . . . . . . . . . . . . . . . . . . 25
1.3 Régime de faible interaction et l’opérateur linéaire . . . . . . . . . . . 26
1.4 Le régime de forte interaction et l’approximation de Thomas-Fermi . 27
1.5 Condensat en rotation . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.5.1 État de l’art . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.5.2 Premier résultat : Non existence de vortex dans la zone de
faible densité d’un condensat . . . . . . . . . . . . . . . . . . . 31
1.6 Réduction de dimension . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.7 L’énergie de Gross-Pitaevskii à deux composants . . . . . . . . . . . . 33
1.7.1 Régime de forte répulsion et faible interaction et
problèmes de partitions optimales . . . . . . . . . . . . . . . . 34
1.7.2 Deuxième résultat : Ségrégation et brisure de symétrie . . . . 36
1.7.3 Le régime de forte répulsion et forte interaction et
problèmes de transition de phase . . . . . . . . . . . . . . . . 37
1.7.4 Troisième résultat : Γ-convergence de l’énergie εEε . . . . . . . 40
1.8 L’opérateur de Schrödinger magnétique
périodique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.8.1 L’opérateur de Harper . . . . . . . . . . . . . . . . . . . . . . 43
1.8.2 Quatrième résultat : Le spectre de l’opérateur de
Schrödinger magnétique avec potentiel de kagome . . . . . . . 46
2 Non existence of vortices 49

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.1.1 Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.1.2 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.1.3 Main ideas of the proof . . . . . . . . . . . . . . . . . . . . . . 53
2.2 Properties of auxiliary functions . . . . . . . . . . . . . . . . . . . . . 55
2.3 Splitting the energy . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4 Proofs of Theorems 2.2 and 2.3 . . . . . . . . . . . . . . . . . . . . . 63
13
14 TABLE DES MATIÈRES
3 Segregation and symmetry breaking 67

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.2 The segregation limit . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.3 Properties of fully segregated components . . . . . . . . . . . . . . . 81
3.4 The non interacting limit . . . . . . . . . . . . . . . . . . . . . . . . . 85
4 A phase transition problem 91

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.1.1 Links with related problems . . . . . . . . . . . . . . . . . . . 96
4.1.2 Main ideas in the proof . . . . . . . . . . . . . . . . . . . . . . 96
4.1.3 To go further . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2 Estimates for ηε . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.3 Rewriting the energy . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4 Upper bound inequality . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.5 Lower bound inequality and compactness . . . . . . . . . . . . . . . . 112
4.5.1 Lower bound on lines . . . . . . . . . . . . . . . . . . . . . . . 112
4.5.2 The slicing method . . . . . . . . . . . . . . . . . . . . . . . . 115
4.5.3 Remark about a lower bound for vε in the transition zone . . . 118
4.6 Proof of the Γ-convergence for ε (Eε (·) − Eε (ηε )) . . . . . . . . . . . . 120
4.6.1 Proof of the compactness and the lower bound inequality in
Theorem 4.1 : . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.6.2 Proof of the upper bound inequality in Theorem 4.1 : . . . . . 120
4.6.3 Proof of Corollary 4.2 : . . . . . . . . . . . . . . . . . . . . . . 124
4.6.4 Proof of Corollary 4.3 . . . . . . . . . . . . . . . . . . . . . . 124
4.7 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5 The magnetic Schrödinger operator 131

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.2 The kagome lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.2.1 Definition and labeling . . . . . . . . . . . . . . . . . . . . . . 141
5.2.2 Study of the symmetries . . . . . . . . . . . . . . . . . . . . . 143
5.2.3 Construction of kagome potentials . . . . . . . . . . . . . . . . 145
5.3 The Schrödinger magnetic operator on L2 (R2 ) . . . . . . . . . . . . . 153
5.3.1 The Schrödinger magnetic operator . . . . . . . . . . . . . . . 153
5.3.2 Quantization of G(Γ) . . . . . . . . . . . . . . . . . . . . . . . 153
5.3.3 Construction of a basis of the space attached to the low lying
spectrum of Ph,A,V . . . . . . . . . . . . . . . . . . . . . . . . 161
5.4 The interaction matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.4.2 The nearest neighbors interaction matrix . . . . . . . . . . . . 169
5.4.3 The representation using discrete translations . . . . . . . . . 172
5.4.4 Representation using pseudo differential operators . . . . . . . 176
5.4.5 Hofstadter’s butterfly for the kagome lattice . . . . . . . . . . 180
Table des figures
1.1 Condensats unidimensionnels . . . . . . . . . . . . . . . . . . . . . . 22

1.2 Papillon de Hofstadter . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3 Réseau de kagome . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.4 Papillon dans le cas d’un réseau triangulaire . . . . . . . . . . . . . . 47
1.5 Papillon dans le cas d’un réseau hexagonale . . . . . . . . . . . . . . 48
5.1 Square, triangular, hexagonal and kagome lattices . . . . . . . . . . . 133

5.2 Hofstadter’s butterfly . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.3 Hofstadter’s butterfly for the triangular lattice . . . . . . . . . . . . . 137
5.4 Hofstadter’s butterfly for the hexagonal lattice . . . . . . . . . . . . . 139
5.5 Kagome pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.6 Vectors ν1 , ν3 , ν5 , 2ν1 and 2ν2 . . . . . . . . . . . . . . . . . . . . . . 142
5.7 Labeling of the kagome lattice . . . . . . . . . . . . . . . . . . . . . . 142
5.8 Vectors ν1 , ν3 , ν5 , µ1 , µ3 and µ5 . . . . . . . . . . . . . . . . . . . . . 146
5.9 Potentials Vj and V from (5.2.33) . . . . . . . . . . . . . . . . . . . . 150
5.10 Kagome potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.11 Kagome potential for p ∈ {2, 10, 30, 50} . . . . . . . . . . . . . . . . 152
5.12 Spectrum of Wγ for some rational values of γ . . . . . . . . . . . . . 182
5.13 Hofstadter’s butterfly for the kagome lattice . . . . . . . . . . . . . . 183
15
16 TABLE DES FIGURES
Chapitre 1
Introduction
1.1 La physique des condensats de Bose-Einstein

1.1.1 La condensation de Bose-Einstein
En 1924, Bose envoie un article à Einstein, où il dérive la loi de Planck pour le
rayonnement du corps noir à partir de raisonnements statistiques. Einstein s’inté-
resse beaucoup au travail de Bose, traduit l’article en allemand et le soumet pour
sa publication ([39]). Einstein applique ensuite les idées de Bose à un gaz idéal
([73]), et prédit que si la densité spatiale n des particules du gaz est plus grande
qu’une certaine valeur critique nc , alors une fraction macroscopique des particules
se “condense” dans l’état d’énergie le plus bas possible pour une seule particule.
Une formulation équivalente de cette idée revient à comparer la distance moyenne
1
particules du gaz d = n− 3 avec la longueur d’onde de de Broglie thermique
entre les √
λT = h/ 2πmkB T , où kB et h sont respectivement la constante de Boltzmann et
de Plank et m est la masse des particules. Le phénomène de condensation arrive
lorsque d est du même ordre de grandeur que λT . Si le gaz est formé de fermions, le
principe d’exclusion de Pauli exclut la possibilité d’avoir plusieurs particules dans
le même état. En revanche, avec un gaz de bosons, on peut s’attendre à ce qu’une
grande partie des particules soit dans le même état.
Pendant plusieurs décennies, la prédiction d’Einstein demeure une prévision théori-

que, car aucun système physique connu ne présente les propriétés nécessaires à la
condensation. Une condition essentielle est d’avoir un gaz très dilué, de façon à s’ap-
procher de l’hypothèse d’Einstein d’un gaz parfait, c’est-à-dire, sans interactions.
Ceci implique que le gaz doit être très froid et donne un aspect important des ex-
périences. Le gaz ne peut pas être contenu dans une boîte, car dans ce cas, toutes
les particules se colleraient aux parois de la boîte. Il est donc nécessaire de confiner
les particules grâce à un champ magnétique inhomogène, qui maintient le gaz en
lévitation au milieu d’un vide très poussé.
À partir des années 60, grâce au développement des lasers, il est possible de ma-
nipuler des atomes en utilisant leur interaction avec les photons. Dans les années
80, des techniques sont mises au point pour refroidir des atomes avec des faisceaux
lumineux. Les travaux développés par Chu, Cohen-Tannoudji et Phillips dans ce
17
18 CHAPITRE 1. INTRODUCTION
domaine sont récompensés par le prix Nobel de Physique en 1997 ([54], [56], [137],
[149]). Les techniques de refroidissement des atomes par des faisceaux lumineux et
par des champs magnétiques permettent l’obtention expérimentale de la conden-
sation pour la première fois en 1995. Le groupe de Cornell et Wieman à Boulder
observe un condensat d’atomes de rubidium ([24]), et le groupe de Ketterle au MIT
un condensat d’atomes de sodium ([68]). Ces trois physiciens sont couronnés par le
prix Nobel de Physique en 2001 ([61], [113], [138]) pour leurs travaux.
L’expérience se déroule de la manière suivante : les atomes sont d’abord refroidis

grâce à des lasers. Puis, les atomes sont placés au centre d’un piège magnétique. En-
fin, le nuage atomique piégé est refroidi à nouveau grâce à une technique “d’évapora-
tion”, qui consiste à laisser les atomes les plus énergétiques s’échapper du piège.
Cette procédure permet d’atteindre des températures de l’ordre du micro Kelvin et
les atomes se condensent.
L’observation du condensat est faite grâce à un éclairage lumineux, dont l’absorption

ou le déphasage sont mesurés. On renvoie au compte rendu [35] de Bloch, Dalibard
et Zwerger sur les expériences et la théorie des gaz dilués ultra froids, ainsi qu’aux
livres de Pethick et Smith [148] et de Pitaevskii et Stringari [150].
1.1.2 L’approximation de Hartree et l’énergie de

Gross-Pitaevskii
L’approximation de Hartree permet de représenter un condensat de Bose-Einstein
à partir d’une fonction complexe u, appelée la fonction d’onde du condensat. Le
condensat est alors décrit par l’énergie de Gross-Pitaevskii
Z
1 g
Eg (u) = |∇u|2 (x) + |x|2 |u|2 (x) + |u|4 (x) dx . (1.1.1)
2 R3 2
On présent ici la justification heuristique de cette approximation (voir [9] ou [64]
pour plus de détails). Considérons un gaz de N bosons piégés dans un potentiel V (x)
et interagissant entre eux par l’intermédiaire d’un potentiel à deux corps w(|x1 −x2 |).
Dans le formalisme de la mécanique quantique, les énergies du système sont données
par les valeurs propres de l’Hamiltonien
N
X p2i 1 X
H= + V (xi ) + w(|xi − xj |) , (1.1.2)
i=1
2m 2 1≤i<j≤N
où pi = −i~∇xi est l’opérateur d’impulsion pour la i-ième particule. On commence

par chercher l’état fondamental de l’hamiltonien H dans l’espace des fonctions de
Hartree :
Ψ(x1 , . . . , xN ) = u(x1 ) · · · u(xN ) u ∈ L2 (R3 ; C) . (1.1.3)

Ces fonctions satisfont par construction au principe de symétrisation des bosons.
2
R quantité N |u(x)| s’interprète comme la densité du gaz au point x. En calculant
La
R3
Ψ(HΨ) on a :
1.1. LA PHYSIQUE DES CONDENSATS DE BOSE-EINSTEIN 19
~2
Z Z
2 2
Ψ(HΨ) = N |∇u(x)| + V (x)|u(x)| dx (1.1.4)
R3 R3 2m
Z
N (N − 1)
Z
2 2
+ |u(x)| |u(y)| w(x − y)dy dx .
2 R3
On trouve N contributions identiques provenant du terme d’énergie cinétique et du

terme d’énergie potentiel extérieur, et N (N − 1) contributions provenant du terme
d’énergie d’interaction entre les particules.
Dans un gaz très froid, la longueur d’onde thermique des particules est grande en
comparaison à la porté du potentiel d’interaction w. Les particules ne sont pas
sensibles à la forme particulière du potentiel w et le seul paramètre significatif est
la longueur de diffusion a, qui dans l’approximation de Born (voir par exemple [64])
s’écrit
Z
m
a' w(x) dx . (1.1.5)
4π~2 R3
R
Les minimiseursRde l’énergie R3 Ψ(HΨ) varientR lentement à l’échelle de w, et on
peut remplacer |u(y)|2 w(x − y)dy par |u(x)|2 w(y)dy dans le dernier terme de
(1.1.4). Compte tenu de (1.1.5), l’énergie totale du système dans l’approximation de
Hartree s’écrit
~2 2π~2 a(N − 1)
Z
2 2 4
E(u, N ) = N |∇u(x)| + V (x)|u(x)| + |u(x)| dx .
R3 2m m
R
La minimisation de cette énergie sous la contrainte R3 |u|2 = 1 donne l’équation
~2 4π~2 a
− ∆u(x) + V (x)u(x) + (N − 1) |u(x)|2 u(x) = µu(x) ,
2m m
où µ est le multiplicateur de Lagrange. En multipliant cette équation par ū et en
intégrant sur tout l’espace, on trouve que le multiplicateur de Lagrange est égal au
potentiel chimique du système, c’est-à-dire, à l’énergie nécessaire pour ajouter une
particule dans le système :
µ = E(u, N ) − E(u, N − 1) .
Lorsque le potentiel de piégeage est un potentiel harmonique V (x) = mω 2 |x|2 /2, on
se ramène à une énergie sans dimension en prenant respectivement
1/2
~
= ~ω et d= (1.1.6)
mω
comme échelles d’énergie et de distance. En effet, en définissant le paramètre sans
dimension g = 8π(N − 1)a/d, l’énergie s’écrit
E(u, N ) = N Eg (u) ,
où Eg est l’équation de Gross-Pitaesvkii définie dans (1.1.1). Le nombre de particules

étant très grand dans les expériences (entre 104 et 107 ) il est courant de remplacer
N − 1 par N dans la définition de g :
8πN a
g≈ . (1.1.7)
d
Le paramètre g joue un rôle essentiel pour le comportement du condensat. Il re-
présente les interactions entre les particules du condensat, lesquelles peuvent être
attractives ou répulsives. Dans ce dernier cas, la taille du condensat augmente avec le
nombre de particules. Une caractéristique importante est que la valeur de a (et donc
aussi la valeur de g) peut être modulée dans les expériences en modifiant légèrement
le potentiel d’interaction entre les particules par des champs externes. Ceci permet
d’obtenir une grande plage des valeurs pour g, ce qui justifie les régimes asympto-
tiques g très grand ou g très petit étudiés dans les modèles théoriques de cette thèse.
La démarche heuristique entreprise ci-dessus pour déduire l’énergie de Gross-Pitaevskii

à partir de l’Hamiltonien des N particules peut être formalisée rigoureusement. La
longueur de diffusion (voir [118]) peut être définie par

φ(r)
a = lim r − 0 ,
r→∞ φ (r)
où φ est la solution de l’équation
~ 00 1
− φ (r) + w(r)φ(r) = 0 avec φ(0) = 0 .
2m 4
Notons eqm (N, a) l’énergie fondamentale dans l’espace de fonctions (1.1.3) de l’Ha-
miltonien (1.1.2) avec ~ = 1 = 2m et V (x) = |x|2 , et ea l’infinimum de 2E8πa dans
L2 (R3 ). Lieb, Seiringer et Yngvason prouvent dans [119] la relation
eqm (N, aN −1 )
lim = ea ,
N →∞ N
dans le cas d’un gaz de bosons dilué avec d’interactions répulsives entre les particules.
1.1.3 Configurations expérimentales

Condensats en rotation
Une autre configuration expérimentale consiste à faire tourner les atomes piégés se-
lon un axe. Cette configuration peut être obtenue en superposant au potentiel de
confinement un potentiel anisotrope en rotation, créé grâce à un faisceau laser mo-
dulé ou un champ magnétique en rotation.
Lorsque le condensat est soumis par exemple à une rotation d’axe z et de vitesse Ω,
il acquiert une moment cinétique −ΩLz non nul dans la même direction, où
Lz = −i~(x1 ∂x2 − x2 ∂x1 ) .

L’approximation de Hartree conduit alors à une fonctionnelle d’énergie similaire à
(1.2.1), à laquelle on doit ajouter un terme d’énergie venant du moment cinétique :
Z
1 g
Eg,Ω (u) = |∇u|2 + ω 2 |x|2 |u|2 + |u(x)|4 − 2 Ω ū · (Lz u) dx , (1.1.8)
2 R3 2
où u·v = 21 (uv̄ + ūv). La dérivation rigoureuse de la fonctionnelle de Gross-Pitaevskii
dans les cas d’un condensat piégé, en rotation selon une direction et avec un terme
d’interaction répulsif, a été faite dans [120]. Il est courant d’additionner et soustraire
un terme à l’énergie pour l’écrire sous la forme
Z
1 2 2 2 2 2 2 2 2 1
~ × ~r u| + (ω − Ω )|r| |u| + ω |x3 | |u| + g|u| dx ,
4
Eg,Ω (u) = |∇u − iΩ
2 R3 2
(1.1.9)
~
où Ω = Ω(0, 0, 1) et ~r = (−x2 , x1 , 0).
Pour obtenir un condensat en rotation, la vitesse Ω doit être inférieure à la fréquence

ω du piège magnétique. Sinon, le potentiel de piégeage ne parvient pas à confiner
les particules et elles sont expulsées du piège par la rotation.
Lorsque Ω = 0, les minimiseurs de la fonctionnelle d’énergie (1.1.8) ont une phase

constante et ne s’annulent pas. Si Ω > 0, les minimiseurs de Eg,Ω acquièrent une
phase complexe grâce au dernier terme dans (1.1.8). En écrivant u = |u|eiS , ∇S est
identifiée à une vitesse et on associe une circulation au condensat par ∇ × ∇S, qui
est nulle si la densité du condensat ne s’annule pas. Au dessus d’une certaine valeur
de la vitesse de rotation, le condensat développe des singularités où u s’annule et
autour desquelles il y a une circulation non nulle. Ces singularités forment des lignes
et lorsqu’on regarde une section transversale du condensat dans le plan xy, les points
où u s’annule sont appelés “vortex”. Plus la vitesse de rotation est importante, plus
des vortex sont observés. Lorsque Ω est très proche ω, le condensat présente de vortex
dans tout l’espace, lesquels s’organisent dans un réseau triangulaire appelé “réseau
d’Abrikosov” ([2]), et dont la taille moyenne devient comparable à leur distance
mutuelle.
Réduction de dimension et périodicité
L’utilisation de lasers ou de potentiels magnétiques permet de confiner les atomes

dans des configurations de dimensions réduites ou périodiques. Par exemple, un
réseau optique 2D dans le plan xy est généré par la superposition de deux ondes
optiques stationnaires avec des polarisations orthogonales et est placé dans le plan
perpendiculaire à un piège magnétique. Dans cette configuration, la fonctionnelle de
Gross-Pitaesvkii du condensat s’écrit
Z
1 2 1 2 2 2 2 1 4
E (u) = |∇u| + 2 Wper (x1 , x2 )|u| + ω |x3 | |u| + g|u| dx .
2 R3 2
Pour des réseaux optiques d’une profondeur assez grande (0 < 1), les atomes
sont contraints de bouger uniquement dans la direction transversale. Les atomes sont
alors confinés dans des petits “tubes” unidimensionnels, ressemblant à des cigares
et placés perpendiculairement dans les minima du réseau 2D (voir Figure 1.1). De
façon similaire, un potentiel périodique unidimensionnel est généré en superposant

deux faisceaux lasers se propageant en directions opposées. Une onde stationnaire
périodique est formée et les atomes sont piégés dans ses minima, formant des couches
de condensats approximativement bidimensionnels.
Figure 1.1 – Répresentation schématique des condensats unidimensionnels (gris)

dans un réseau périodique 2D (couleur).
Condensats de Bose Einstein à deux composants
Des condensats avec deux types de particules sont obtenus expérimentalement en

utilisant deux types d’atomes différents ([75], [128], [132], [163]), deux isotopes dis-
tincts du même atome ([146]), ou encore un même isotope avec différentes structures
hyperfines ([69], [85], [124], [127], [136]). Dans l’approximation de Hartree, le conden-
sat de Bose-Einstein à deux composants est décrit par deux fonctions d’onde com-
plexes u1 et u2 , dont les modules au carré représentent respectivement les densités
du premier et du deuxième composant (voir [148], chapitre 12). L’énergie du système
est donnée par la somme des énergies de Gross-Pitaevskii de chaque composant, et
d’une énergie provenant des interactions entre les particules de deux composants.
Cette dernière ne dépend que des densités des condensats. Lorsque les particules
des deux composants ont la même masse (m1 = m2 ) et qu’elles sont piégées dans le
même potentiel magnétique de fréquence ω, l’énergie sans dimensions 1 du condensat
à deux composants est donnée par
Z
g12
E(u1 , u2 ) = Eg1 (u1 ) + Eg2 (u2 ) + |u1 (x)|2 |u2 (x)|2 dx ,
2 R3
où Eg est l’énergie de Gross-Pitaevskii définie dans (1.1.1). Le paramètre g12 cor-

respond à la force d’interaction entre les particules du premier composant et du
deuxième composant et le paramètre g1 (respectivement g2 ) correspond à la force
1. On choisit ici et d (définies dans (1.1.6)) comme échelles d’énergie et de distance.
d’interaction entre les particules du premier composant (respectivement du deuxième

composant). On signale que si le potentiel de piégeage est magnétique, alors le confi-
nement des particules dépend de leur structure hyperfine, donc les potentiels pour
chaque composante ne sont pas forcement égaux. Le cas général où m1 6= m2 et
V1 6= V2 a été étudié dans [126].
Dans les expériences physiques ([85], [128], [146]) et lors des simulations numériques
([109], [110], [126], [142]) on observe la séparation des supports des composants.
L’étude de ce phénomène, qu’on appelle “ségrégation spatiale”, est une des motiva-
tions principales de cette thèse. La ségrégation des composants s’étudie en fonction
du paramètre
2
g12
Γ=1− . (1.1.10)
g1 g2
Quand Γ est positif et grand, chaque composant est asymptotiquement dans le ré-
gime de Thomas-Fermi d’un condensat à un composant (voir la Section 1.4 plus
loin). Ainsi, u1 et u2 sont à symétrie radiale et proches d’une parabole inversée. Les
supports sont des disques concentriques avec des rayons distincts.
Quand Γ devient négatif, les composants perdent leur symétrie radiale et leurs sup-
ports prennent des formes proches de demi-disques. Quand Γ est négatif de grand
module, les composants sont complètement ségrégés, à l’exception d’une fine couche
d’interface. Si g1 > g2 , le bord du support de u1 à une courbure positive et le bord du
support de u2 à une courbure négative. Si g1 = g2 les supports sont des demi-disques.
Les paramètres g1 et g2 peuvent être variés en modifiant la quantité des particules

utilisées dans le condensat ([128], [146]). Le paramètre g12 peut être varié en modi-
fiant la structures hyperfine des particules à l’aide des champs externes ([163]). Il
est ainsi possible de manipuler les condensats à deux composants pour obtenir dif-
férentes configurations spatiales. Ceci justifie les limites théoriques considérées dans
cette thèse pour g1 , g2 et g12 .
Les simulations numériques d’un condensat à deux composants mis en rotation

mettent en évidence des défauts topologiques. Comme dans le cas à un seul compo-
sant, des vortex sont présents, mais ces vortex peuvent s’ordonner dans des réseaux
triangulaires ou carrés. Dans [14], Aftalion, Mason et Wei dérivent une énergie pour
les vortex permettant de discriminer entre les deux types de réseau. D’autres phé-
nomènes apparaissent aussi : les “coreless vortex” (vortex dans un composant ac-
compagnés de pics dans l’autre composant), les “vortex sheets” (lignes de vortex), et
les “giant skyrmions” (un anneau des singularités). Une classification de ces défauts
topologiques en fonction de la vitesse de rotation Ω et du paramètre Γ a été faite
par Aftalion et Mason dans [126]. Dans cette thèse on ne traite pas le cas de deux
composants avec rotation.
1.1.4 Magnétisme artificiel avec des condensats de

Bose-Einstein
La simulation de phénomènes quantiques avec des systèmes physiques classiques

constitue actuellement un sujet de recherche très important en physique expérimenta-
le et théorique (voir le compte rendu [35] de Bloch, Dalibard et Zwerger à ce sujet).
Un des objectifs majeurs est l’obtention de magnétisme artificiel à partir des atomes
neutres, dans le but de récréer des phénomènes magnétiques des systèmes quan-
tiques, tel que l’effet Hall quantique ([134], chapitre 3).
Cet but a été atteint avec par exemple des condensats de Bose-Einstein en rotation.
En effet, en regardant l’écriture (1.1.9) de l’énergie d’un condensat en rotation à
vitesse Ω, on aperçoit que les deux premiers termes correspondent à l’énergie d’une
particule de masse m = 1 et charge q dans un champ magnétique q B ~ = 2Ω.~ Ceci
crée un champ magnétique artificiel dans le référentiel tournant et a été largement
utilisé (voir [65] et les références là-dedans). Or, des restrictions techniques limitent
la vitesse de rotation et cette approche ne permet pas d’atteindre les intensités de
champs magnétiques nécessaires à l’observation de l’effet Hall quantique.
Comme on l’a vu précédemment, la structure interne des atomes peut être modi-
fiée en utilisant des faisceaux lumineux. En particulier, il est possible d’ajuster les
interactions entre les particules d’un gaz en fonction de leur position spatiale. Sous
certaines conditions bien contrôlées ([79], [102], [107]) ceci entraîne qu’une particule
à deux états quantiques se déplaçant adiabatiquement à travers un contour fermé
acquerra une phase. Cette phase est appelée “géométrique” (ou de Berry) car elle
ne dépend que de la trajectoire suivie. Par analogie avec l’effet d’Aharonov-Bohm
([134], chapitre 1), cette phase manifeste le présence d’un champ magnétique ex-
terne. Cette idée a été implémentée avec des condensats de Bose-Einstein ([122]).
On mentionne ici que Panati, Spohn et Teufel développent dans [143], [144] et [145]
une théorie adiabatique pour l’étude de l’approximation Born-Oppenheimer. Ré-
cemment, Aftalion et Nier étudient rigoureusement une approximation adiabatique
pour une particule à deux états quantiques dans un faisceau laser ([15]), suivant un
modèle introduit dans [83].
Dans [99], Hou propose un schéma théorique pour l’obtention de magnétisme arti-
ficiel avec un condensat bidimensionnel placé dans un réseau optique à symétrie de
kagome. L’auteur y étudie numériquement le spectre du système obtenu, en fonction
de l’intensité du champ magnétique artificiel. Compte tenu de la périodicité du ré-
seau de kagome, un dessin similaire au célèbre papillon de Hofstadter ([98] et Figure
5.2) y est présenté.
1.2. LA FONCTIONNELLE DE GROSS-PITAEVSKII 25
Figure 1.2 – Papillon de Hofstadter. Cette figure représente le spectre d’énergies

d’un particule chargée dans un réseau carré 2D soumis à un champ magnétique
constante dans la direction perpendiculaire. Le paramètre γ est relié aux flux du
champ magnétique à travers une cellule de périodicité du réseau carré.
1.2 La fonctionnelle de Gross-Pitaevskii

Ce mémoire de thèse porte sur des modèles mathématiques pour des condensats à
deux dimensions. Le principal modèle utilisé est la fonctionnelle d’énergie de Gross-
Pitaevskii, qui associé à u dans un espace convenable son énergie
Z
1 2 1 2 g 4
Eg (u) = |∇u(x)| + V (x)|u(x)| + |u(x)| dx . (1.2.1)
R2 2 2 4
On rappelle que le premier terme correspond à l’énergie cinétique du condensat,
le deuxième à son énergie potentielle (ici on suppose V continu et positif) et le
troisième à l’énergie d’interaction entre ses particules. On se restreint au cas où les
particules du condensat interagissent répulsivement. On suppose donc
g ≥ 0.
La minimisation de Eg s’effectue dans l’espace
Z
1 2 2
H = u ∈ H (R ; C) ; V (x)|u(x)| dx < ∞ , (1.2.2)
R2
et sous la contrainte de masse

Z
|u(x)|2 dx = 1 . (1.2.3)
R2
L’injection de H dans L2 (R2 ; C) est compacte lorsque V est un potentiel localement

borné tel que inf |x|≥R V (x) → +∞ quand R → ∞ (voir [119]). Ceci assure l’existence
de minimiseurs de Eg dans H sous la contrainte (1.2.3). Dans ce mémoire de thèse

on travaille particulièrement avec des potentiels satisfaisant
1 p
∃C > 0 , ∃p ≥ 2 , V (x) ≥
|x| ∀|x| > C . (1.2.4)
C
Le potentiel de piégeage le plus souvent considéré est le potentiel magnétique
V (x) = ω12 x21 + ω22 x22 . (1.2.5)

Si u est un minimiseur de Eg dans H sous la contrainte (1.2.3), alors il satisfait
l’équation de Gross-Pitaevskii
− ∆u + V (x)u + g|u|2 u = λg u , (1.2.6)

où λg est le multiplicateur de Lagrange. D’après (1.2.3), on trouve la valeur de λg
en multipliant (1.2.6) par ū puis en intégrant :
Z
|∇u(x)|2 + V (x)|u(x)|2 + g|u(x)|4 dx .

λg = (1.2.7)
R2
L’espace de minimisation H contient des fonctions à valeurs complexes, mais on peut

se restreindre aux fonctions réelles strictement positives. En effet, d’après l’inégalité
diamagnétique ([117], section 7.21), si u est un minimiseur, alors |u| l’est aussi. À
partir de (1.2.6), la régularité elliptique de l’équation et un argument de bootstrap
impliquent que u est une fonction régulière ([80], chapitre 8). Donc, en appliquant le
principe du maximum fort ([80], chapitre 3) on a |u| > 0. La structure de l’équation
(1.2.6) et la contrainte de masse (1.2.3) impliquent qu’il y a unicité des paires de
solutions (u, λg ), avec u > 0. On trouve alors que tout minimiseur de Eg dans H
sous la contrainte (1.2.3) est de la forme
ug = ηg eiθ2
où θ2 ∈ R et ηg est la seule solution strictement positive de (1.2.6).
On étudie naturellement l’énergie Eg quand g est un petit ou un grand paramètre,

correspondant respectivement aux régimes de faible ou forte interaction entre les
particules du condensat. On renvoie au [4] pour plus de détails.
1.3 Régime de faible interaction et l’opérateur li-

néaire
Dans le régime de faible interaction, l’énergie cinétique et l’énergie du potentiel de
piégeage sont grandes devant l’énergie d’interaction entre les particules. On peut
alors négliger cette dernière, et lorsqu’on considère le potentiel de piégeage magné-
tique dans (1.2.5), on se retrouve avec l’énergie d’un oscillateur harmonique en deux
dimensions
Z
1 2 1 2 2 2 2 2
Ehar (u) = |∇u(x)| + (ω1 x1 + ω2 x2 )|u(x)| dx . (1.3.1)
R2 2 2
1.4. LE RÉGIME DE FORTE INTERACTION ET L’APPROXIMATION DE THOMAS-FERMI27
On s’attend à que l’état de base du condensat soit proche de la première fonction

propre de l’opérateur linéaire
1 1 1
Hhar (u) = − ∆ + ω12 x21 + ω22 x22 , (1.3.2)
2 2 2
donnée par la gaussienne
√
ω1 ω2 /2 − 1 (ω1 x21 +ω2 x22 )
1
ϕ(x) = e 2 .
π
En effet, notant ug le minimiseur de Eg sur H et eg son énergie, la projection de ug
sur ϕ permet d’estimer aisément
kug − cϕkL2 (R2 ) ≤ Cg /2

1
avec |c| = 1. De même, l’énergie du condensat est proche de la première valeur

propre de Hhar :
√
ω1 + ω2 ω1 ω2
g + Og→0 (g /2 ) .
3
eg = +
2 8π
1.4 Le régime de forte interaction et l’approxima-

tion de Thomas-Fermi
Dans le régime de forte interaction, l’énergie d’interaction entre les particules est
grande devant l’énergie cinétique et l’énergie du potentiel de piégeage. Ce cas est
appelé régime de Thomas-Fermi, et a été largement étudié dans la littérature ([4],
[5], [11], [16], [77], [100], [101], [108]). En négligeant l’énergie cinétique dans (1.2.1),
on est conduit à minimiser
Z
TF 1 2 g 4
E (u) = V (x)|u(x)| + |u(x)| dx .
R2 2 4
Lorsque V est donné par (1.2.5), cette fonctionnelle admet un unique minimiseur
(voir [116]) parmi les fonctions u ∈ L2 (R2 ) ∩ L4 (R2 ) satisfaisant (1.2.3). Le minimi-
seur satisfait l’équation d’Euler-Lagrange
V (x)u(x) + g|u(x)|2 u(x) = λu(x) ,

qui donne la solution 2
1/2
λ − V (x)
|u| = (1.4.1)
g +
avec λ déterminé par (1.2.3). Cette fonction n’appartient pas à l’espace de minimisa-
tion H (voir plus loin), mais elle donne une première approximation dans le régime
de Thomas-Fermi. Lorsqu’on considère un potentiel de piégeage magnétique, l’état
fondamental d’un condensat est donc proche d’une parabole inversée. Pour obtenir
des estimations plus précises, on introduit un petit paramètre
2. Pour une fonction a réelle on note a+ = max(a, 0) et −a− = min(a, 0).
ε = g − /2 > 0 ,
1
et on effectue un changement d’échelle en posant
ũ(x) = ε− /2 u(ε− /2 x) .
1 1
(1.4.2)
Remarquons que la régime de Thomas-Fermi correspond à la limite quand ε → 0 et
que le changement d’échelle préserve la contrainte de masse (1.2.3) :
Z Z
2
|ũ(x)| dx = |u(x)|2 dx .
R2 R2
L’énergie de Gross-Pitaevskii (1.2.1) vérifie alors Eg (u) = εẼε−2 (ũ) où Ẽε−2 est
l’énergie de Gross-Pitaevskii d’un condensat dans un piège magnétique de fréquence 3
ωε−1 . Ceci amène à considérer de façon générale la fonctionnelle d’énergie
Z
1 2 1 2 1 4
Eε (u) = |∇u(x)| + 2 V (x)|u(x)| + 2 |u(x)| dx (1.4.3)
R2 2 2ε 4ε
pour un potentiel de piégeage quelconque satisfaisant (1.2.4).
Au regard de l’approximation grossière trouvée dans (1.4.1), on s’attend à ce que le

minimiseur ait une faible densité hors de
D = x ∈ R2 ; V (x) ≤ λ .

(1.4.4)
Le compact D est appelé la “zone de haute densité du condensat”. Le comportement
du minimiseur de Eε est très important dans l’étude des condensats en rotation et des
condensats à plusieurs composants. Il est donc fondamental de disposer d’estimations
très précises sur le minimiseur strictement positif de Eε dans H, c’est-à-dire
Eε (ηε ) = inf Eε (u) , ηε > 0 , (1.4.5)

u∈H , kukL2 (R2 ) =1
lequel satisfait l’équation d’Euler-Lagrange

1 1
− ∆ηε + 2
V (x)ηε + 2 ηε3 = λε ηε . (1.4.6)
ε ε
Pour estimer ηε on définit
a(x) = λ − V (x) ,
et en utilisant (1.2.3) on réécrit
Z
1
Eε (ηε ) = Eε1 (ηε ) + 2 λ− 2
(a+ (x)) dx ,
2ε R2
où
Z
1 1 2 1
Eε1 (η) = |∇η|2 + |η|2
− a + + a− |η|2 .
2 R2 2ε2 ε2
3. Ici on a utilisé l’homogénéité du piège magnétique.
1.4. LE RÉGIME DE FORTE INTERACTION ET L’APPROXIMATION DE THOMAS-FERMI29
On prévoit que lorsque ε est petit, ηε soit proche du “profile de Thomas-Fermi”

√
ρ= a+ . (1.4.7)
Pour majorer l’énergie on voudrait utiliser ρ comme fonction test, mais la dérivée de
ρ explose au bord de D et ρ n’est pas dans H 1 (D). On construit alors une fonction
test η̂ε en recollant ρ à zéro de façon linéaire près de ∂D. On trouve alors qu’il existe
C > 0 tel que pour ε suffisamment petit
Eε1 (ηε ) ≤ Eε1 (η̂ε ) ≤ C| ln ε| .

En utilisant l’analogue de (1.2.7) pour Eε , l’inégalité ci-dessus permet d’estimer le
multiplicateur de Lagrange, lequel vérifie
|ε2 λε − λ| ≤ ε| ln ε|2 .
On résume ici des estimations disponibles pour ηε en fonction de ρ, lorsque le po-
tentiel est magnétique et donné par (1.2.5).
Théorème 1.1. Il existe ε0 > 0, c1 , c2 , C > 0, α ∈ (1/2, 3/5) et γ ∈ (1/2, 3/4), et pour
tout K ⊂⊂ D, CK > 0
√
kηε − ρkC 1 (K) ≤ CK ε2 | ln ε| , (1.4.8)
√
|ηε (x) − ρ(x)| ≤ C εγ α
for x ∈ B(0 , λ − c1 ε ) , (1.4.9)
1
ε− /3 (λ−|x|)
ηε (x) ≤ C ε /6 ec2 for x ∈ R2 \D
1
(1.4.10)
pour ε ∈ (0, ε0 ).
L’estimation (1.4.9) près du bord de D, a été obtenue par Gallo et Pelinovsky dans
[77] (voir aussi [76]) pour le minimiseur de Eε sans contrainte sur la norme L2 . L’idée
principale est que près de ∂D, ηε est approché par
1 − |x|2

1/3
ε ν ,
ε2/3
où ν est l’unique solution l’équation de Painlevé-II
00
4ν (y) + yν(y) − ν 3 (y) = 0 , y ∈ R,
telle que
ν(y) ∼ y /2
1
quand y → +∞ et ν(y) → 0 quand y → −∞ .
On remarque que récemment des estimations similaires on été établies par Karali
et Sourdis dans [108] pour le cas d’un potentiel V non radial satisfaisant ∂r V > 0
près de D. Dans l’article [12], qui sera présenté au Chapitre 2, nous avons prouvé
que lorsque le potentiel de piégeage est à symétrie radiale, et qu’il est une fonction
croissante de la variable radiale, le minimiseur de Eε est une fonction décroissante
de la variable radiale près du bord.
On signale qu’un autre modèle souvent utilisé, est celui d’un condensat à deux
dimensions avec condition de Dirichlet homogène sur le bord de la zone de haute
densité. On est alors amené à étudier l’énergie de Gross-Pitaevskii (voir (1.4.3))
Z
1 2 1 2 1 4
Êε (u) = |∇u(x)| + 2 V (x)|u(x)| + 2 |u(x)| dx
2 D ε 2ε
R
dans H01 (D) avec contraint D |u(x)|2 dx = 1.
On n’étudie pas ce type de problème ici et on renvoie à [4] ou [5] pour des estima-
tions de ηε dans ce cas.
1.5 Condensat en rotation

1.5.1 État de l’art
La fonctionnelle d’énergie de Gross-Pitaevskii décrivant un condensat en rotation
a été largement étudié dans la littérature (voir exemple [6], [7], [8], [16], [27], [62]
parmi beaucoup d’autres). Lorsqu’on considère un condensat à deux dimensions et
piégé dans un piège magnétique de fréquence ω > 0, la fonctionnelle est donnée par
Z
1 2 1 2 2 2 g 4 ⊥

Eg,Ω̃ (u) = |∇u| + ω |x| |u| + |u| − Ω̃ x · (iu, ∇u) dx , (1.5.1)
R2 2 2 2
où Ω̃ > 0 est la vitesse de rotation, x⊥ = (−x2 , x1 ) et (iu, ∇u) = iu∇ū − iū∇u.

Cette fonctionnelle d’énergie a été largement étudiée dans la littérature ([6], [7], [8],
[16], [62]). On a vu précédemment que la fréquence du piégeage magnétique doit
être plus forte que la force centrifuge pour que les particules restent confinées dans
le piège. En effet, l’existence des minimiseurs est assurée lorsque Ω̃ < ω. Si Ω̃ > ω il
est possible de construire des suites de fonctions test dont l’énergie tend vers −∞,
et par conséquence, il n’y a pas de minimiseurs.
L’étude de l’énergie de Gross-Pitaevskii dans le cas de rotation rapide (Ω̃ ∼ ω) est

faite dans l’espace propre associé au premier niveau de Landau. Ce régime n’est pas
traité dans ce mémoire et on renvoie à [4], [6] ou [8] pour plus des détails. Dans le
cas de la rotation lente, l’énergie est étudiée dans le régime de Thomas-Fermi décrit
dans la Section 1.4. En faisant le même changement d’échelle qu’en (1.4.2) et en
introduisant Ω = ε−1 Ω̃, on est amené à considérer, pour un potentiel de piégeage
quelconque, l’énergie
Z
1 2 1 4 1 2 ⊥

Eε,Ω (u) = |∇u| + 2 |u| + 2 V (x)|u| − Ω x · (iu, ∇u) dx . (1.5.2)
R2 2 4ε 2ε
La minimisation de cette énergie s’effectue dans le espace H défini dans (1.2.2) et

sous la contrainte (1.2.3). Un outil important, introduit par Lassoued et Mironescu
dans [114], est la réécriture de Eε,Ω comme la somme de l’énergie du profile du
1.5. CONDENSAT EN ROTATION 31
condensat et l’énergie de sa phase complexe. En définissant v = u/ηε , l’énergie se

réécrit
Eε,Ω (u) = Eε (ηε ) + Fε (v) (1.5.3)

où
ηε2 ηε4
Z
2 2 2 2 ⊥

Fε (v) = |∇v| + 2 (|v| − 1) − ηε Ω x · (iv, ∇v) dx (1.5.4)
R2 2 4ε
et ηε est défini dans (1.4.5).
Dans le cas où on minimise Eε,Ω dans D, avec condition de Dirichlet sur ∂D et avec un
potentiel magnétique, Aftalion et Du prouvent dans [10] que la vitesse critique pour
l’existence de vortex est de l’ordre de | ln ε|. Ils calculent aussi la vitesse critique
pour l’existence de d vortex dans D. Dans le cas où le potentiel correspond à la
superposition d’un potentiel magnétique et d’un faisceau laser :
V (x) = (1 − b)x2 + k|x|4 (1.5.5)

avec b > 1 + (3k 2 /4)1/3 , la vitesse critique est déterminée par Aftalion, Alama et
Bronsard dans [5]. Dans ce cas la zone de haute densité est un anneau. Ces auteurs
prouvent que la vitesse critique pour l’existence de vortex dans l’anneau est
Ωc = ω0 | ln ε| + ω̃0 ln | ln ε| (1.5.6)
avec ω0 , ω̃0 > 0.
Le cas sur tout R2 est traité par Ignat et Millot dans [100] et [101]. Ces auteurs
déterminent la vitesse critique pour l’existence de d vortex à l’intérieur de la zone
de haute densité, ainsi que l’énergie de la configuration des vortex. Ils considèrent
le cas d’un potentiel magnétique du type (1.2.5), avec ω1 = 1 et ω2 ∈ (0, 1]. La zone
de haute densité est l’ellipse {(x1 , x2 ) ∈ R2 ; x21 + Λx22 ≤ λ2 }, et la vitesse critique
pour l’existence d’un vortex à l’intérieur de l’ellipse est donnée par (1.5.6), avec
ω0 = (ω22 + 1)λ−2 .
Ces travaux s’inspirent des techniques développées par Bethuel, Brezis et Hélein
(voir le livre [34]) pour l’étude des vortex dans le modèle du type Ginzburg-Landau,
et par S. Sandier et S. Serfaty (voir le livre [156]) dans le cas avec champ magnétique.
1.5.2 Premier résultat : Non existence de vortex dans la zone

de faible densité d’un condensat
Lors d’une rencontre entre physiciens et mathématiciens, Lev Pitaevskii pose la
question suivante : est-ce que les vortex apparaissent d’abord dans la zone de haute
densité D, ou au contraire, ils sont créés à l’extérieur de D et rentrent dans D quand
la vitesse de rotation dépasse Ωc ? Les travaux mentionnés précédemment établissent
la vitesse critique pour l’existence de vortex dans la zone de haute densité et ne
permettent pas de déterminer si d’autres vortex existent en dehors de cette zone et
jusqu’à l’infini. Dans un travail en collaboration avec Aftalion et Jerrard, qui fait
l’objet du Chapitre 2 de ce mémoire de thèse, nous donnons la réponse à la question

posée par Pitaevskii. Numériquement et expérimentalement, la faible densité du
condensat en dehors de D rend difficile la détection des vortex. L’existence d’une
preuve mathématique apporte ainsi une information importante. Nous prouvons
qu’un condensat en rotation à un vitesse sous-critique n’a aucun vortex dans R2 .
Nous établissons ce résultat en montrant que pour ε assez petit, le minimiseur de
Eε,Ω est égal (modulo un terme de phase constante) au minimiseur ηε de Eε , qui
lui, ne s’annule pas (voir (1.4.5)). Plus surprenant, notre résultat montre que le
condensat ne ressent pas la rotation jusqu’à que celle-ci dépasse la valeur critique
Ωc ! Plus précisément, le théorème principal est le suivant :
Théorème 1.2. Soit V un potentiel à symétrie radiale, satisfaisant (1.2.4) et tel

que ∂r V > 0 dans un voisinage ouvert de ∂D. Soit uε un minimiseur de Eε,Ω (·).
Il existe ε0 , ω0 , ω1 > 0 tels que, si 0 < ε < ε0 et Ω ≤ ω0 | ln ε| − ω1 ln | ln ε|, alors
uε = eiθ2 ηε dans R2 , pour une constante θ2 réelle.
Perpectives
Le Théorème 1.2 ci-dessus n’est établi que dans le cas où le condensat est dans
un piège à symétrie radiale. Ceci est une limitation technique et le résultat devrait
rester valide lorsqu’on considère un piège non radial, mais satisfaisant les conditions
de croissance à l’infini (1.2.4) et près du bord de la zone de haute densité (∂ν V > 0
où ν est le vecteur normal à ∂D). Notre preuve utilise des fonctions auxiliaires dont
on ne connait des estimations que dans le cas radial. Ces fonctions auxiliaires sont
définies à partir de l’état fondamental ηε définit dans (1.4.5). Les estimations pour
ηε récemment prouvées par Karali et Sourdis dans [108] dans le cas non radial,
pourraient ouvrir une voie à la généralisation du Théorème 1.2.
1.6 Réduction de dimension

Ce mémoire de thèse ne comprend que des résultats pour des condensats à deux
dimensions, sans considérer le comportement du gaz dans la troisième dimension.
Une justification rigoureuse de cette approximation a été fournie par Aftalion et
Helffer dans [11]. Dans leur travail, ces auteurs étudient un condensat en dimension
3, en rotation à une vitesse Ω dans un piège magnétique, auquel on superpose un
potentiel périodique W dans la direction z de profondeur 1/, où > 0 est un petit
paramètre. En considérant un condensat périodique dans la direction z, Aftalion et
Helffer déterminent des conditions sur les paramètres Ω, et g pour que l’énergie
du condensat se découple de deux façons : soit comme la somme de l’énergie d’un
condensat en dimension 2 dans le plan xy et l’énergie d’un oscillateur harmonique
dans la direction z ; soit comme la somme de l’énergie d’un condensat en dimension
1 dans la direction z et de l’énergie d’un oscillateur harmonique dans le plan xy. Par
exemple, en supposant que la fréquence du piège magnétique dans le plan xy est
beaucoup plus forte que dans la direction z (ω⊥ ωz ), la réduction au modèle d’un
condensat à deux dimensions dans le plan xy est justifiée par l’estimation suivante :
e = λz + eg,ω⊥ ,Ω (1 + o(1)→0 ) ,
1.7. L’ÉNERGIE DE GROSS-PITAEVSKII À DEUX COMPOSANTS 33
où e est l’énergie du condensat à trois dimensions, λz est la première valeur de

d2
l’oscillateur harmonique Hz = − 12 dz 2 + W (z) dans une cellule de périodicité de
W , et eg,ω⊥ ,Ω est le minimum de l’énergie Eg̃,Ω dans (1.5.1) avec g̃ = (2π)−1 gω⊥ .
Ces preuves utilisent des techniques de l’analyse semi-classique, et des techniques
de l’analyse non linéaire utilisées par exemple dans l’étude de Thomas-Fermi (voir
Section 1.4 plus haut).
1.7 L’énergie de Gross-Pitaevskii à deux composants

Dans cette thèse, on étudie la fonctionnelle d’énergie de Gross-Pitaevskii pour un
condensat à deux composants, qui dans sa version sans dimension, et en notant
g = (g1 , g2 , g12 ), est donnée par
Z
1
Eg (u, v) = Eg1 (u1 ) + Eg2 (u2 ) + g12 |u1 |2 |u2 |2 dx ,
2 R2
où Egj (j = 1, 2) est l’énergie d’un condensat donnée dans (1.2.1). On rappelle que
les paramètres g1 , g2 et g12 correspondent aux interactions entre les différent types
des particules. On se restreint aux cas d’égalité entre les masses des particules, et
on considère toujours que les deux composants sont dans le même piège magnétique
à symétrie radiale :
V (x) = ω 2 |x|2 , ω ≥ 0.
On étudie le cas d’interactions répulsives entre toutes les particules, ce qui revient
à considérer
g1 ≥ 0 , g2 ≥ 0 et g12 ≥ 0 .
On suppose aussi que le nombre total de particules du système est conservé, et
qu’il ne peut pas y avoir d’échange de particules entre les deux composants. La
minimisation de Eg s’effectue alors dans l’espace des paires de fonctions (u1 , u2 ) ∈
H × H avec les contraintes de masse
Z Z
|u1 | dx = α1 et |u2 | dx = α2 , (1.7.1)
R2 R2
où α1 + α2 est fixé. Les paires minimisantes de Eg satisfont le système de deux
équations de Gross-Pitaevskii couplées

−∆u1 + V u1 + g1 |u1 |2 u1 + g12 |u2 |2 u1 = λ1,g u1
, (1.7.2)
−∆u2 + V u2 + g2 |u2 |2 u2 + g12 |u1 |2 u2 = λ2,g u2
où λ1,g et λ2,g sont les multiplicateurs de Lagrange associés à (1.7.1).
Grâce à des arguments similaires à ceux de la Section 1.2 dans le cas d’un seul
composant, les paires minimisantes vérifient
u1 = eiθ1 |u1 | , |u1 | > 0 et θ1 ∈ R ,

u1 = eiθ2 |u2 | , |u2 | > 0 et θ2 ∈ R .
Si (u1 , u2 ) est une paire minimisante et R est une rotation du plan, alors (u1 ◦ R, u2 ◦
R) est encore une paire minimisante, mais contrairement au cas d’un seul compo-
sant, la symétrie radiale de chaque composant ne découle pas de la symétrie radiale
du potentiel de piégeage.
Comme on l’a vu dans la Section 1.1.3, lorsque le paramètre g12 est très grand, le
condensat à deux composantes présente ségrégation spatiale, et une brisure de sy-
métrie apparaît. La ségrégation a fait l’objet de nombreux travaux mathématiques
(voir la Section 1.7.1 suivante), mais aucun travail rigoureux n’avait été mené pour
comprendre la brisure de symétrie. Une des motivations principales de cette thèse
est l’étude de cette brisure de symétrie. On étudie ces sujets dans deux régimes
particuliers.
Le premier régime est celui de forte répulsion entre les particules des deux compo-
sants, avec faible répulsion entre les particules de chaque composant. Ceci correspond
à l’étude des minimiseurs de Eg dans les limites
g12 → +∞ , g1 → 0 et g2 → 0 . (1.7.3)
Le deuxième régime étudié est celui de forte répulsion entre les particules des deux
composants, avec forte répulsion entre les particules de chaque composant. En par-
ticulier, on suppose que les forces d’interaction entre les particules d’un même com-
posante sont égales. Ceci correspond à l’étude des minimiseurs de Eg dans les limites
g12 → +∞ et g1 = g2 → +∞ . (1.7.4)
La différence fondamentale entre ces deux cas vient des valeurs de l’énergie. Dans
les deux cas, la ségrégation des composants est assez forte pour maintenir le dernier
terme de Eg borné. Alors, dans le régime (1.7.3) l’énergie reste bornée parce que
g1 et g2 sont petits, et les support de chaque composant peuvent s’étendre jusqu’à
l’infini. Par contre, dans le régime (1.7.4), l’énergie explose car les composants sont
contraints à partager la zone de haute densité D, qui elle, est bornée. On explique
dans la suite le cadre mathématique utilisé dans chaque régime.
1.7.1 Régime de forte répulsion et faible interaction et

problèmes de partitions optimales
L’étude des systèmes d’équations elliptiques du type (1.7.2), avec un paramètre de
couplage très grand (dans notre cas g12 1), est appelée “problème de ségrégation”.
Ces systèmes sont largement étudiés dans la littérature dans le cas où les fonctions
u1 et u2 sont définies dans un domaine borné de Rn ([31], [51], [53], [140], [165]).
Dans la limite g12 très grand, leurs solutions présentent une ségrégation spatiale, et
le problème limite devient un système d’équations elliptiques non couplées. De façon
générale, chaque composant vérifie son équation limite juste dans son support, et sa
dérivée présente un saut à travers l’interface tout en restant bornée. La différence
des deux composants est une fonction régulière. Dans notre cas, on est amené à
étudier le système

−∆u1 + V u1 + g1 |u1 |2 u1 = λ1 u1 dans {|u1 | > 0}
, (1.7.5)
−∆u2 + V u2 + g2 |u2 |2 u2 = λ2 u2 dans {|u2 | > 0}
où λ1 (respectivement λ2 ) est la limite de λ1,g (respectivement λ2,g ) quand g12 → ∞
(voir (1.7.2)).
L’existence et régularité des solutions de systèmes du type (1.7.2) sont étudiées par
Caffarelli et Lin, dans le cas d’un domaine borné U, avec V = 0, g1 = g2 = 0
et λ1 = λ2 = 0. Dans [51], ils prouvent que les composants limites sont harmo-
niques à l’intérieur de leurs supports et Lipschitziens dans U tout entier. Ces auteurs
montrent aussi que l’interface entre les composants est d’intérieur vide. La conver-
gence uniforme vers des composants limites est prouvée par Wei et Weth ([165]) dans
le cas V = 0. Ces résultats sont améliorés par Noris, Tavares, Terracini et Verzini
dans [140], qui prouvent des bornes Hölder en fonction de g12 , ainsi que la continuité
Lipschitz des composants limites. Ces auteurs utilisent les formules de monotonicité
d’Almgren et Alt, Caffarelli et Friedman ([20]). On renvoie aussi à [49], [53], [58] et
[59] pour d’autres résultats dans le même esprit.
Un autre aspect important dans l’étude des systèmes du type (1.7.2), est le compor-
tement asymptotique des solutions près de l’interface. Ceci est étudié par Berestycki,
Lin, Wei et Zhao dans [31], qui dérivent le profile des composants près de l’interface
(voir aussi [33]). Dans le cas d’un domaine borné en dimension 1, ils prouvent la
continuité Lipschitz uniforme des composant. Via une technique d’éclatement (blow-
up), ils dérivent aussi les équations près de l’interface. Dans le cas de la dimension
2, ils prouvent que sous certaines hypothèses, les composants limites doivent être
unidimensionnels près de l’interface.
On signale aussi que des systèmes proches de (1.7.2), mais avec V = 0, g1 < 0 et
g2 < 0 sont étudiés par Dancer, Wei et Weth dans [67], [164] et [166]. Ce cas est très
différent car l’énergie Eg n’est pas nécessairement bornée inférieurement.
La ségrégation totale de composants conduit à un problème de frontière libre appelé

“problème de partitions optimales” (voir par exemple [47], [57] et [60]). Dans notre
cas, chaque composant doit minimiser son énergie Eg , et les supports doivent être
disjoints. Dans [53], des résultats numériques sont donnés pour le problème d’un
condensat avec k composants totalement ségrégés.
À la limite g1 = g2 = 0, les équations limites dans (1.7.5) deviennent linéaires et

on est amené à considérer de partitions optimales dites “spectrales”. Il s’agit, en
considérant les premières valeurs propres λ(U1 ) et λ(U2 ) des réalisations de Dirichlet
de Hhar dans deux ouverts disjoints U1 et U2 de R2 , de minimiser le maximum ou
la somme de λ(U1 ) et λ(U2 ). Dans le cas de la somme, U1 et U2 correspond aux
supports de chaque composant. Ce type de problème fait l’objet de nombreuses
recherches ces dernières années. Dans [50], l’existence de partition optimales et la
régularité des interfaces sont prouvées, dans le cas de la minimisation de la somme.
Le cas de la minimisation du maximum pour un opérateur de Schrödinger est étudié
par Helffer, Hoffmann-Ostenhof et Terracini dans , [89], [90] et [91]. On notera
que dans le cas de la somme, le problème est délicat même pour le cas de deux
composants. Par exemple, la détermination d’une partition optimale pour la somme
associée au Laplacien sur le disque est un problème ouvert. On renvoie à [87] pour
un compte rendu des résultats mathématiques, ainsi qu’à [36] et [41] pour des études
numériques.
1.7.2 Deuxième résultat : Ségrégation et brisure de symétrie

Dans l’article [153], qui fait l’objet du Chapitre 3, j’étudie la ségrégation et la brisure
de symétrie d’un condensat à deux composants dans un piège magnétique, dans le
cas où les deux composants ont la même masse, et dans le régime (1.7.3). La ca-
ractéristique principale du système est que les composants se séparent, tandis que
chaque composant se trouve dans le régime linéaire de la Section 1.3.
J’étudie d’abord la ségrégation des composants dans la limite g12 → +∞. Je montre
la convergence localement uniforme vers une paire limite, et que chaque composant
de la paire limite est solution d’une équation de Gross-Pitaevskii dans son propre
support. J’obtiens aussi l’énergie limite du système. Voici le premier résultat :
Théorème 1.3. Soit (u1,gn , u2,gn )n∈N une suite de minimiseurs de Egn dans H × H,
satisfaisant (1.7.1) avec α1 = α2 = 1, et avec g n = (g1n , g2n , g12
n
) ∈ [0, c0 ]2 × R+ tel
que
lim g1n = g1 , lim g2n = g2 et n

lim g12 = ∞. (1.7.6)
n→∞ n→∞ n→∞
Alors, il existe une paire (u1,∞ , u2,∞ ) ∈ H × H tel que u1,∞ · u2,∞ = 0 et
(i) (u1,gn , u2,gn ) converge (à l’extraction d’une sous suite près) vers (u1,∞ , u2,∞ ),
faiblement dans HV1 (R2 ) × HV1 (R2 ) et dans Cloc
0
(R2 ) × Cloc
0
(R2 ).
(ii) (u1,∞ , u2,∞ ) minimise Eg1 ,g2 ,0 parmi les paires (u1 , u2 ) ∈ H × H, satisfaisant
(1.7.1) avec α1 = α2 = 1 et telles que u1 · u2 = 0.
(ii) La paire (u1,∞ , u2,∞ ) est solution du système

 −∆u1 + V u1 + g1 |u1 |21 u1 = λ∞ u 1 dans {|u1,∞ | > 0}
−∆u2 + V u2 + g2 |u2 |2 u2 = µ∞ u2 dans {|u2,∞ | > 0} , (1.7.7)
u1 · u2 = 0 dans R2

où λ∞ et µ∞ sont les limites respectives des multiplicateurs de Lagrange λgn

et µgn , associés à (u1,gn , u2,gn ) par (1.7.1).
Pour prouver que les composants d’une paire minimisante n’ont pas la symétrie
du potentiel (brisure de symétrie), j’étudie ensuite le problème limite lorsque g1 =
g2 = 0. Il s’agit d’un problème des partitions optimales spectrales pour l’oscillateur
harmonique sur tout R2 . Je prouve que les partitions optimales sont constituées de
deux demi-espaces complémentaires, et que la paire limite optimale correspondante
est déterminée par une deuxième fonction propre de l’oscillateur harmonique :
Théorème 1.4. Soit V (x) = ω 2 |x|2 . Soit (u1,0 , u2,0 ) un minimiseur de E0,0,0 parmi
les paires (u1 , u2 ) ∈ H × H, satisfaisant (1.7.1) avec α1 = α2 = 1 et telles que
u1 · u2 = 0. Il existe ν ∈ S 1 tel que (i) le support de u1,0 est {x · ν > 0} et le support

de u2,0 est {x · ν < 0} ; et
(ii) u1,0 and u2,0 sont les parties positives et négatives de la deuxième fonction propre
de l’oscillateur harmonique −∆+ω 2 |x|2 dans R2 ayant ces deux demi-espaces comme
ensembles nodaux.
La preuve de la brisure de symétrie découle alors facilement de la convergence loca-

lement uniforme et de l’étude du problème des partitions optimales :
Corollaire 1.5. Soit V (x) = ω 2 |x|2 . Il existe des constantes g0 > 0 et G > 0
telles que si (ug , vg ) est un minimiseur de Eg dans H × H, satisfaisant (1.7.1) avec
α1 = α2 = 1, et
max{g1 , g2 } ≤ g0 et g12 ≥ G ,
alors ug et vg ne sont pas radiales.
Perpectives
Le Théorème 1.3 sur la ségrégation des composants n’est établi que dans le cas d’éga-
lité des masses α1 = α2 = 1 et dans le cas d’un potentiel harmonique V (x) = ω 2 |x|2 .
Les techniques utilisées dans la preuve devraient se généraliser facilement au cas où
α1 6= α2 et V est un potentiel de piégeage satisfaisant (1.2.4).
Le Corollaire 1.5 sur la brisure de symétrie dépend fortement de la forme du potentiel

harmonique. Il découle du Théorème 1.4, dont la preuve repose sur un changement
de variable qui transforme le problème initial en un problème de minimisation de
la somme de deux quotients de Rayleigh gaussien dans le plan (voir la Section 3.4
au Chapitre 3). La clé de la preuve est la propriété minimale des demi-espaces par
rapport au quotient de Rayleigh gaussien. Cette structure du problème limite est
particulière au potentiel harmonique. Il est donc peu probable que ma preuve de la
brisure de symétrie puisse s’étendre à d’autres potentiels radiaux. Il serait cependant
intéressant de développer d’autres techniques pour étudier la brisure de symétrie.
Les études sur les problèmes de partitions optimales spectrales ([87]) pourraient
apporter des idées dans cette direction.
1.7.3 Le régime de forte répulsion et forte interaction et

problèmes de transition de phase
Dans ce régime, g1 et g2 sont très grands, donc chaque composant se trouve dans le
régime de Thomas-Fermi pour un condensat à un composant. De façon analogue à
la Section 1.4, on introduit un petit paramètre ε > 0, on écrit
1
g1 = g2 = et g12 = gε ,
ε2
et on étudie la limite quand ε → 0 de l’énergie
Z
1
Eε (u1 , u2 ) = Eε (u1 ) + Eε (u2 ) + gε |u1 |2 |u2 |2 dx ,
2 R2
où Eε est l’énergie de Gross-Pitaevskii donnée dans (1.4.3). La condition Γ → −∞

(voir 1.1.10) s’écrit alors
ε2 gε → +∞ quand ε → 0. (1.7.8)
De façon analogue au régime de Thomas-Fermi pour un seul composant (voir la
Section 1.4), la somme des densités des deux composants est négligeable hors de
D. En plus, comme les deux composants sont en forte répulsion, ils présentent une
ségrégation spatiale, avec deux support disjoints D1 et D2 , tels que D1 ∪ D2 = D.
Chaque composant est alors proche dans son support d’un morceau de la parabole
inversée définie en (1.4.7), et développe un saut de hauteur ρ à travers le bord de son
support. On est donc confronté à un problème de transition de phase, où le premier
composant passe de 0 à ρ à travers le bord d’un ensemble A ⊂ D, tandis que le
deuxième composant passe de ρ à 0.
Dans cette thèse on étudie rigoureusement ce phénomène, et on donne une caractéri-

sation de l’ensemble A. On présente ici une brève description du modèle utilisé. Sui-
vant une technique issue du modèle σ non linéaire (voir la section 4.2.2. dans [109]),
on introduit une représentation du condensat à deux composants en fonction de sa
densité totale ṽ = |u1 |2 +|u2 |2 et du vecteur de spin ϕ̃ = (|u1 |2 −|u2 |2 )/(|u1 |2 +|u2 |2 ).
L’énergie Eε s’écrit en fonction de ṽ et ϕ̃, et elle comporte un terme d’énergie ciné-
tique de la forme ṽ|∇ϕ̃|/(1− ϕ̃2 ). Pour se débarrasser du dénominateur, on considère
plutôt ϕ = arccos ϕ̃, ce qui revient à définir
!
ϕ |u1 | + i|u2 |
:= Arg p , (1.7.9)
2 |u1 |2 + |u2 |2
où Arg est la valeur principale de l’argument complexe.
On s’inspire du changement de variables de Lassoued et Mironescu dans (1.5.3) pour

séparer l’énergie provenant de l’état de base d’un condensat à un composant. On
définit
p
|u1 |2 + |u2 |2
v= , (1.7.10)
ηε
où ηε est le minimiseur de Eε dans H sous la contrainte (1.2.3) (voir la Section 1.4).
On parvient alors à l’identité suivante pour Eε : si (u1 , u2 ) ∈ H2 est tel que |u1 |2 +
|u2 |2 > 0, alors
Eε (u1 , u2 ) = Eε (ηε ) + Fε (v, ϕ) , (1.7.11)

où F(v, ϕ) = Fε (v) + Gε (v, ϕ), avec
Z
1 2 2 1 4 2 2
Fε (v) = η |∇v| + 2 ηε {1 − v } dx , (1.7.12)
2 R2 ε 2ε
Z
1
ηε2 v 2 |∇ϕ|2 + ηε4 v 4 g̃ε {1 − cos2 (ϕ)} dx .

Gε (v, ϕ) = (1.7.13)
8 R2
De plus, si (u1 , u2 ) satisfait les contraintes dans (1.7.1), alors (v, ϕ) satisfait
Z Z
ηε2 v 2 dx = α1 + α2 et ηε2 v 2 cos ϕ dx = α1 − α2 . (1.7.14)
R2 R2
Dans les deux énergies (1.7.12) et (1.7.13) on trouve la somme d’un terme d’énergie
cinétique et d’une énergie associée à un potentiel à deux puits. Il s’agit donc d’éner-
gies du type “Cahn-Hilliard”, correspondant à la transition continue d’un fluide entre
deux phases.
L’étude des énergies du type “Cahn-Hilliard” est faite dans le cadre de la théorie
de la Γ-convergence. La Γ-convergence est une notion de convergence variationnelle,
introduite dans les années 70 par De Giorgi et Franzoni ([81]). On présente ici l’une
des définitions possibles de la Γ-convergence, et on renvoie au livre de Braides [42]
pour une étude détaillée, ainsi qu’à l’article d’Alberti [17] pour une brève exposition
du sujet.
Definition 1.6. Soit Tn une suite de fonctions définies sur un espace métrique X
et à valeurs dans R. On dit que Tn Γ-converge vers T : X → R, si :
(1) Pour toute suite {xn } ⊂ X telle que |Tn (xn )| est borné, il existe x et une sous-
suite xnk → x dans X, tels que
lim inf Tnk (xnk ) ≥ T (x) .

k→∞
(2) Pour tout x ∈ X, il existe une suite xn → x telle que
lim sup Tn (xn ) ≤ T (x) .

n→∞
La Γ-convergence sert par exemple à dériver l’énergie limite d’un système physique
dans un régime particulier, ou inversement, à étudier les minimiseurs d’un problème
d’une fonctionnelle d’énergie donnée, en regardant des suites de minimiseurs de
fonctionnelles approchées. La Γ-convergence possède les propriétés suivantes :
Proposition 1.7.
(1) Une Γ-limite T est semi continue inférieurement.
(2) Si Tn Γ-converge vers T et S est continue, alors (Tn +S) Γ-converge vers (T +S).
(3) Si Tn Γ-converge vers T et xn minimise Tn dans X, alors tout point d’accumu-

lation de {xn } minimise T dans X.
Revenons à nos énergies Fε et Gε définies dans (1.7.12) et (1.7.13). Si ηε ≡ 1,

alors εFε corresponde à la fonctionnelle de Modica-Mortola ([130]). Dans ce cas, εFε
(définie dans un domaine borné U) Γ-converge vers le problème de minimisation du
périmètre :
σ̂ inf H1 (Sv) , (1.7.15)

v∈BV (U ; {0,1})
R1
où σ̂ = 2−1/2 −1
(1 − t2 ) dt.
Ici BV (U ; {0, 1}) est l’ensemble de fonctions à variation bornée à valeurs dans
{0, 1}, Sv est l’ensemble de singularités essentielles de v et H1 désigne la mesure de
Hausdorff. On renvoie à [21], [74] et [82] pour les concepts de théorie de la mesure
géométrique. Ce résultat est conjecturé par Di Giorgi en 1977 et prouvé peu après
par Modica et Mortola ([131]). Le lien avec l’énergie de Cahn-Hilliard est fait plus
tard en 1984 par Modica ([130]).
Le cas général d’une fonctionnelle d’énergie du type

Z
1
F̂ε (v) = f (ε∇v, v, x) dx
ε U
est étudié par Bouchitté dans [40] (voir aussi [3]). Ce cas corresponde à une énergie
de Cahn-Hilliard avec des poids non uniformes dans l’espace. En généralisant les
technique introduites par Modica ([130]), Bouchitté prouve la Γ-convergence de F̂ε
vers un problème de minimisation du périmètre, avec un poids dépendant de l’en-
veloppe concave de f .
À la différence des cas étudiés par Modica-Mortola et Bouchitté, notre énergie Fε =

Fε + Gε dépend de deux variables v et ϕ. L’énergie pour la première variable est
Fε et fait intervenir ηε . L’énergie pour la deuxième variable est Gε et fait intervenir
ηε v. On trouve en particulier que v multiplie la dérivée de ϕ. Cette caractéristique
est présente dans la fonctionnelle d’Ambrosio-Tortorelli, définie dans [22] (voir aussi
[23]) pour approcher la fonctionnelle de Mumford-Shah ([135]). Elle a l’effet de
forcer v à s’annuler dans la zone où la dérivée de ϕ est très grande. Ce phénomène
apparaît aussi dans l’étude des couches épitaxiales, où une mince couche de matériel
développe des fissures. Ces fissures correspondent aux endroits où v tend vers zéro et
la dérivée de ϕ est très grande ([37], [38]). Dans la Γ-limite, le phénomène décrit ci-
dessus conduit à l’étude d’un problème de minimisation du périmètre. Notre résultat
principal dans le régime (1.7.4) (voir le Théorème 1.11 plus loin) est dans le même
esprit.
1.7.4 Troisième résultat : Γ-convergence de l’énergie εEε

L’étude, en collaboration avec A. Aftalion, d’un condensat à deux composants dans
un piège magnétique et dans le régime (1.7.4) fait l’objet du Chapitre 4. Les caracté-
ristiques principales du système sont que les composants présentent une ségrégation
spatiale, et que chacun se trouve dans le régime de Thomas-Fermi décrit dans la
Section 1.4. Ceci conduit à analyser un problème de transition de phase pour les
deux composants, que nous étudions en réécrivant l’énergie sous la forme (1.7.11).
Nous prouvons d’abord que cette réécriture est bien valide pour les minimiseurs de
Eε . Plus précisément nous montrons :
Proposition 1.8. Soit (u1 , u2 ) une paire minimisante de Eε dans H × H. Alors,
|u1 |2 + |u2 |2 > 0. De plus, si on définit (v, ϕ) par (1.7.9) et (1.7.10) on a
Eε (u1 , u2 ) = Eε (ηε ) + Fε (v, ϕ) .
Nous établissons ensuite une borne supérieure et une borne inférieure pour εFε , ainsi
que la compacité des suites d’énergie uniformément bornée.
Proposition 1.9. (Borne supérieure pour εFε )

Soit ϕ ∈ BVloc (D ; {0, π}). Il existe une suite (vε , ϕε ) ∈ Lip(R2 ; (0, 1] × [0, π]),
convergeant quand ε → 0 vers (1, ϕ) dans L1loc (D) × L1loc (D), telle que
lim sup εFε (vε , ϕε ) ≤ F(ϕ) .

ε→0
Proposition 1.10. (Borne inférieure et compacité pour εFε )

Soit (vε , ϕε ) ∈ Liploc (R2 ; (0, +∞) × [0, π]) telle que
sup kvε kL∞ (K) ≤ CK quel que soit K ⊂⊂ D (1.7.16)

ε>0
et
sup εFε (vε , ϕε ) < +∞ . (1.7.17)
ε>0
Alors, il existe ϕ ∈ BVloc (D; {0, π}) tel que
(vε , ϕε ) → (1, ϕ) dans L1loc (D) × L1loc (D) (1.7.18)
et
lim inf εFε (vε , ϕε ) ≥ F(ϕ) . (1.7.19)

ε→0
En utilisant les bornes pour εFε , nous montrons la Γ-convergence (au niveau des
minimiseurs) de l’énergie εEε vers un problème de minimisation du périmètre avec
un poids en ρ3/2 . On rappelle que ρ est le profile de Thomas-Fermi défini dans (1.4.7).
Nous montrons le théorème suivant :
Théorème 1.11. Soit V (x) = |x|2 et

n Z o
X = ϕ ∈ BVloc (D ; {0, π}) ; ρ cos ϕ dx = α1 − α2 .
D
On définit une fonctionnelle F sur X par

Z
ρ /2 dH1 ,
3
F(ϕ) = σ̂
D ∩ Sϕ
où σ̂ est définit dans (1.7.15).
La fonctionnelle ε(Eε (·, ·) − Eε (ηε )) Γ-converge, par rapport à la métrique L1loc (D) ×
L1loc (D), vers F(·) :
(i) Compacité et borne inférieure. Soit {(u1,ε , u2,ε )}ε>0 une suite de mini-
miseurs de Eε dans H × H satisfaisant (1.7.1) telle que
sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) < +∞ . (1.7.20)

ε>0
Alors, il existe ϕ ∈ X tel que (quitte à extraire une sous suite)
√
L1loc (D) × L1loc (D) , (1.7.21)

(u1,ε , u2,ε ) → ρ 1{ϕ=0} , 1{ϕ=π} dans
et
lim inf ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) ≥ F(ϕ) . (1.7.22)

ε→0
(ii) Borne supérieure. Pour tout ϕ ∈ X, il existe une suite {(u1,ε , u2,ε )}ε>0 ⊂
√
H × H satisfaisant (1.7.1), et convergeant quand ε → 0 vers ρ 1{ϕ=0} , 1{ϕ=π}
dans L1loc (D) × L1loc (D), telle que
lim sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) ≤ F(ϕ) . (1.7.23)

ε→0
Finalement, nous avons prouvé que si α1 et α2 ne sont pas trop petits, alors les
minimiseurs dans X de l’énergie limite F ne sont pas à symétrie radiale. On en
déduit la brisure de symétrie des minimiseurs de Eε pour ε assez petit.
Proposition 1.12. Il existe δ0 ∈ (0, 1/2) tel que si α1 ∈ [δ0 , 1 − δ0 ], alors les mini-
miseurs de F dans X ne sont pas des fonctions radiales.
Corollaire 1.13. Il existe ε0 > 0 et δ0 ∈ (0, 1/2) tels que si α1 ∈ [δ0 , 1 − δ0 ] et

ε ∈ (0, ε0 ), alors les minimiseurs de Eε ne sont pas des fonctions radiales.
Perspectives
1. Minimiseur de l’énergie sur la couche limite
Les résultats numériques montrent deux configurations stables pour les supports
d’un condensat à deux composants dans le régime (1.7.4), avec α1 et α2 ne pas trop
petits. La première est constituée d’un disque et d’un anneau concentriques et la
deuxième de deux sections de disques. On s’attend à qu’il y ait une brisure de sy-
métrie, de façon que les sections des disques constituent la configuration de moindre
énergie.
Dans le Corollaire 1.13, nous montrons que les minimiseurs de l’énergie F ne sont
pas à symétrie radiale, mais nous ne déterminons pas la configuration minimale. Un
problème ouvert très intéressant est la détermination des minimiseurs de F dans X.
2. Convergence des minimiseur de Eε
Un autre problème intéressant est d’améliorer la convergence de u1,ε et u2,ε dans

leurs supports limites, donnés respectivement par {ϕ = 0} et {ϕ = π}. On pourrait
s’attendre à avoir le même type de convergence que celle de ηε vers ρ, c’est-à-dire,
1
Cloc respectivement dans {ϕ = 0} et {ϕ = π}.
1.8. L’OPÉRATEUR DE SCHRÖDINGER MAGNÉTIQUE PÉRIODIQUE 43
3. Autres régimes pour la ségrégation et la brisure de symétrie
Nous étudions la ségrégation spatiale dans le régime (1.7.4) sous les conditions
g1 = g2 → +∞, g12 → +∞ et g1 ≪ g12 (voir (1.7.8)). Dans le cas où les forces
d’interaction entre les deux composants ne sont pas égales, c’est-à-dire g1 6= g2 , un
terme supplémentaire apparaît dans la réécriture (1.7.11) de l’énergie Eε . Il serait
intéressant de déterminer la ségrégation des minimiseurs de l’énergie Eε dans ce cas.
De même, les simulations numériques montrent que la brisure de symétrie a lieu

√
pour g1 g2 < g12 . Il serait intéressant d’étudier la Γ-limite de Eε dans les cas où
g1 = g2 et g12 sont du même ordre. Dans le cas où g1 ≪ g12 , l’énergie des fonc-
tions auxiliaires v et ϕ de découple et la forme du problème limite est déterminée
par Fε (v). Dans le cas où g1 = g2 et g12 sont du même ordre, on s’attend à avoir
un problème limite plus complexe, déterminé par les deux énergies Fε (v) et Gε (v, ϕ).
4. L’énergie Fε
Dans la preuve du Théorème 1.11 nous utilisons des propriétés sur Fε (v, ϕ). Cette
fonctionnelle n’est étudiée que lorsque les fonctions v et ϕ sont définies à partir
des relations (1.7.9) et (1.7.10). Or, l’étude de l’énergie Fε a un intérêt en elle
même. Les bornes inférieures et supérieures obtenues dans les Propositions 1.9 et
1.10 impliquent la Γ-convergence de Fε dans un espace bien choisi, mais nous ne
savons pas définir cet espace pour assurer l’existence de minimiseurs ! En particulier,
nous ne savons pas construire un espace de minimisation pour prendre en compte
les contraintes de masse dans (1.7.14) et la condition v > 0. Ceci n’est pas trivial
car si l’on prend une suite minimisante (vn , ϕn ), la condition vn > 0 ne passe pas
à la limite. En plus, on ne peut assurer la convergence de ϕn que dans la zone où
vn ne s’annule asymptotiquement pas. Les articles de Ambrosio et Tortorelli ([22] et
[23]) sur les fonctionnelles d’approximation de Mumford-Shah pourraient suggérer
des pistes pour étudier plus en détail l’énergie Fε et ses minimiseurs.
1.8 L’opérateur de Schrödinger magnétique

périodique
1.8.1 L’opérateur de Harper
La dernière partie de cette thèse porte sur l’étude, dans un régime semi-classique,
de l’opérateur de Schrödinger magnétique en dimension 2 donné par
Ph,A,V = (hDx1 − A1 (x))2 + (hDx2 − A2 (x))2 + V (x) . (1.8.1)

Ici h > 0 est un “petit” paramètre, Dxj = 1i ∂xj , V est un potentiel périodique associé
à un réseau et A est le potentiel vecteur du champ magnétique également périodique.
Dans le formalisme de la mécanique quantique, un tel opérateur décrit une parti-

cule chargée dans un réseau à deux dimensions, soumis à un champ magnétique
dans la direction transversale. Le spectre de l’opérateur Ph,A,V correspond ainsi aux
énergies admissibles pour la particule. L’obtention expérimentale de magnétisme ar-

tificiel avec des atomes ultra froids (voir la Section (1.1.4)) est une des motivations
de l’étude de cet opérateur. Plus particulièrement, on a voulu comprendre le travail
de Hou dans [99], où le spectre d’un opérateur de Schrödinger du type (1.8.1) est
analysé, en lien avec un condensat de Bose-Einstein dans un réseau de kagome.
Pour le cas A = 0, le spectre de Ph,A,V est constitué de bandes (voir par exemple
le chapitre XIII.16 dans [152]). Le cas général, même en supposant que le champ
magnétique est constant, est très délicat. On est souvenant conduit à l’étude de
modèles limites dans différents régimes asymptotiques. En particulier, l’étude de
modèles discrets joue un rôle essentiel. Dans certains cas, on est conduit à l’étude
d’un opérateur pseudo-différentiel (ou des systèmes) sur R. La réduction aux diffé-
rents modèles dépend entre autres de la structure du réseau. On explique dans la
suite ces modèles, dans la cas le plus simple d’un réseau carré, lequel conduit à
l’opérateur de Harper dont le spectre (décrit en fonction du flux) décrit le célèbre
papillon de Hofstadter.
Un modèle discret est l’opérateur autoadjoint défini sur `2 (Z2 ) par
Lγ = τ1 + τ1∗ + t(τ2 + τ2∗ ) , (1.8.2)

où t > 0 et
(τ1 u)n,m = un−1,m , (τ2 u)n,m = e2iπγn un,m−1 .

Le paramètre γ ci-dessus est égal au flux du champ magnétique à travers une cellule
de périodicité du réseau, divisé par h. Cet opérateur a été appelé par B. Simon
l’opérateur de “presque-Mathieu” (en référence à l’opérateur de Mathieu défini sur
d2
la droite par − dx 2 + cos x ).
En utilisant une théorie partielle de Floquet par rapport à la deuxième variable

([152], chapitre XII.16), on peut réduire l’étude ensembliste du spectre de Lγ en
introduisant la condition de Floquet
τ2 u = eiθ2 u .
On est amené à étudier une famille d’opérateurs Lγ,θ2 sur `2 (Z) et on a
[
σ(Lγ ) = σ(Lγ,θ2 ) ,
θ2 ∈[0,2π]
où σ(L) désigne le spectre de L.
Quand γ est irrationnel le spectre de Lγ,θ2 ne dépend pas de θ2 . Les opérateurs en

question prennent la forme d’un opérateur de Schrödinger discret sur `2 (Z) et sont
définis par
(Lγ,θ2 v)n = vn+1 + vn−1 + Vθ2 (n)vn , (1.8.3)

avec Vθ2 (n) = 2t cos (2πγn + θ2 ) le potentiel. L’opérateur ci-dessus est appelé, lorsque
t = 1, “opérateur de Harper”. En 1976, Hofstadter ([98]) réalise une étude numérique
du spectre de Lγ,θ2 en fonction de γ. Sa démarche suggère une structure fractale du

spectre et conduit au papillon de Hofstadter. Elle consiste en l’étude du spectre de
l’opérateur quand γ = p/q, avec p, q ∈ N premiers entre eux et p < q ≤ 50. Dans ce
cas, le spectre est formé de q bandes distinctes qui ne peuvent se toucher qu’à leurs
extrémités. Le papillon de Hofstadter est obtenu en mettant sur la droite d’abscisse
p/q les bandes du spectre (voir la Figure 1.2). De plus Hofstadter propose des règles
pour déterminer la configuration des bandes en liaison avec le développement en
fraction continue de p/q. Cette configuration laisse deviner la structure de Cantor
de l’opérateur Lγ,θ2 quand γ est irrationnel. Un problème longtemps ouvert, proposé
par Simon et Kac dans les années 80 et appelé “Ten Martinis problem” ([159], pro-
blème 4), était de prouver que pour tout irrationnel γ et t > 0, le spectre de Lγ est
un ensemble de Cantor. Ce problème est résolu en 2009 par Avila et Jitomirskaya,
après beaucoup des résultats partiels commençant par l’article de Bellissard and
Simon en 1982 ([30]). On renvoie à [26] pour une liste complète des références.
Pour calculer numériquement le spectre de Lγ lorsque γ = p/q, on utilise à nouveau

la théorie de Floquet. En introduisant la condition
vn+1 = eiθ1 vn ,
on se ramène à l’étude spectrale des matrices hermitiennes dans Mq (C) définies par
Mp,q,θ1 ,θ2 = eiθ1 K + e−iθ1 K ∗ + eiθ2 Jp,q + e−iθ2 Jp,q

∗
,
où Jp,q = diag(exp (2iπ(j − 1)p/q)) et

1 si j = i + 1 (mod q)
Kij = .
0 sinon
On a alors
[
σ(Lγ ) = σ(Mp,q,θ1 ,θ2 ) ,
θ1 ,θ2 ∈[0,2π]
et il suffit de diagonaliser les matrices Mp,q,θ1 ,θ2 . Les bandes du spectre sont obtenus
en remarquant que
det (Mp,q,θ1 ,θ2 − λIdq ) = fp,q (λ) + (−1)q+1 2(cos qθ1 + cos qθ2 ) , (1.8.4)
où fp,q (λ) est un polynôme de degré q. Cette relation est appelé “formule de Cham-
ber” (voir [94], section 9). Par exemple, on trouve

cos θ2 cos θ1
M1,2,θ1 ,θ2 = 2 .
cos θ1 − cos θ2
et
det (M1,2,θ1 ,θ2 − λ) = λ2 − 4 cos2 θ1 − 4 cos2 θ2 = (λ2 − 4) − 2(cos 2θ1 + cos 2θ2 ) ,
d’où p
λ(θ1 , θ2 ) = ±2 cos2 θ1 + cos2 θ2 .
Pour revenir à l’étude de l’opérateur sur `2 (Z2 ), on peut aussi montrer qu’il est
unitairement équivalent à l’opérateur pseudo-différentiel agissant sur L2 (R) par
f (x + γ) + f (x − γ)
Lγ f (x) = + cos (x) f (x) , (1.8.5)
2
qui correspond à la γ-quantification de Weyl du symbole cos x + cos ξ. On ren-
voie aux notes de Lein dans [115] pour une introduction de la quantification de
Weyl et au livre [19] d’Alinhac et Gérard pour les concepts reliés aux opérateurs
pseudo-différentiels. Dans une série d’articles ([94], [96], [97]) Helffer et Sjöstrand dé-
veloppent des techniques, inspirées par les travaux du physicien Wilkinson ([168]),
pour étudier l’opérateur Lγ . Leur méthode, qui met en jeu une procédure de renor-
malisation associée au développement en fraction continue de γ, n’est développée
dans cette thèse que dans sa première étape.
Le cas d’un réseau triangulaire est étudié par Claro en Wannier ([55]). Ces auteurs
trouvent une structure analogue au cas du réseau carré, et pour une valeur p/q du
flux magnétique, avec p,q premiers entre eux et p < q, le spectre se divise dans q
bandes (voir Figure 1.4). Le cas d’un réseau hexagonal a été étudié rigoureusement
par P. Kerdelhue ([111], [112]) suivant les techniques de Helffer et Sjöstrand (voir la
Figure 1.5 pour le dessin du spectre). On remarque ici que ce dernier cas correspond
exactement au cas d’un électron dans une couche de graphène ([134], chapitre 6)
avec champ magnétique transversal. Ces travaux retrouvent une nouvelle application
compte tenue de l’intérêt suscité par le graphène après le prix Nobel de physique en
2010 ([78], [139], [141]).
1.8.2 Quatrième résultat : Le spectre de l’opérateur de

Schrödinger magnétique avec potentiel de kagome
L’étude du spectre des opérateurs décrivant une particule chargée dans un potentiel
périodique à deux dimensions et soumis à un champ magnétique constant dans
la direction transversale conduit à une observation commune. Si les minima du
potentiel constituent un réseau de Bravais avec n = 1 ou 2 points par cellule de
périodicité, et si le flux magnétique à travers la cellule de périodicité est égale à
2πp/q, avec p, q ∈ N premiers entre eux et p < q, alors dans certains régimes
asymptotiques on observe au bas du spectre une première zone spectrale constituée
de nq bandes exponentiellement petites séparées du reste du spectre. Ceci est observé
et expliqué par de nombreux physiciens ([98], [55]), et démontré pour le cas du réseau
carré ([94], [96], [97]), triangulaire et hexagonale ([111], [112]). L’étude du réseau de
kagome (voir Figure 1.3) est entrepris dans cette thèse et fait l’objet du Chapitre
5. Le réseau de kagome est un réseau de Bravais avec trois points par cellule de
périodicité. Ce cas est étudié par les physiciens Hou ([99]) et Xiao et all. ([169]).
D’abord, j’effectue la réduction rigoureuse de l’opérateur de Schrödinger sur L2 (R2 )
à une famille d’opérateurs sur `2 (C3 ). Ensuite, j’écris un programme sur Matlab
pour calculer les spectres des opérateurs sur `2 (C3 ). Enfin, j’étudie les symétries du
dessin obtenu. Motivée par la réalisation expérimentale des potentiels optiques, je
construis des familles de potentiels ayant les symétries du réseau de kagome, étant à
priori réalisables expérimentalement (car polynômes trigonométriques) et vérifiant
quelques hypothèses des travaux mathématiques précédents.
Figure 1.3 – Réseau de kagome.
Figure 1.4 – Papillon dans le cas d’un réseau triangulaire. Le paramètre γ est
relié aux flux du champ magnétique à travers une cellule de périodicité du réseau
triangulaire.
Figure 1.5 – Papillon dans le cas d’un réseau hexagonale. Le paramètre γ est
relié aux flux du champ magnétique à travers une cellule de périodicité du réseau
hexagonale.
Chapitre 2
Non existence of vortices in the small

density region of a condensate
Amandine Aftalion 1 , Robert L. Jerrard 2 and Jimena Royo-Letelier 3
Article apparu le 15 Avril 2011

dans Journal of Functional Analysis
Abstract
In this paper, we answer a question raised by Lev Pitaevskii and prove that
the ground state of the Gross Pitaevskii energy describing a Bose Einstein
condensate in a rotationally symmetric trap at low rotation does not have
vortices in the low density region. Therefore, the first ground state with vor-
tices has its vortices in the bulk. In fact we prove something stronger, which
is that the ground state for the model at low rotations is equal to the ground
state in a condensate with no rotation. This is obtained by proving that for
small rotational velocities, the ground state is multiple of the ground state
with zero rotation. We rely on sharp bounds of the decay of the wave function
combined with weighted jacobian estimates.
2.1 Introduction
Among the many experiments on Bose Einstein condensates, one consists in rotating
the trap holding the atoms in order to observe a superfluid behaviour : the appea-
rance of quantized vortices [1, 161, 148, 150, 4]. This takes place for sufficiently large
rotational velocities. On the contrary, at low rotation, no vortex is detected in the
bulk of the condensate. The system can be described by a complex valued wave
function minimizing a Gross Pitaevskii type energy. A vortex corresponds to zeroes
1. CNRS et Université Versailles-Saint-Quentin-en-Yvelines, Laboratoire de Mathématiques de
Versailles, CNRS UMR 8100, 45 avenue des États-Unis, 78035 Versailles Cédex, France
2. Dept. of Mathematics University of Toronto, Toronto, Canada M5S2E4.
3. CMAP, Ecole Polytechnique, 91128 Palaiseau cedex, France, and Université Versailles-Saint-
Quentin-en-Yvelines, Laboratoire de Mathématiques de Versailles, CNRS UMR 8100, 45 avenue
des États-Unis, 78035 Versailles Cédex, France.
49
50 CHAPITRE 2. NON EXISTENCE OF VORTICES
of the wave function with phase around it. The density of the condensate is signifi-
cant in a region which is either a disk or an annulus, and gets exponentially small
outside this domain. Vortices are experimentally visible in the bulk of the conden-
sate. A question raised by Lev Pitaevskii is whether for small rotational velocity,
when there are no vortices in the bulk, vortices could exist in the low density region.
For very large rotational velocities, when bulk vortices are arranged on a triangular
lattice, it has been shown [7] that in a simplified model, obtained by formally projec-
ting the Gross-Pitaevsky energy onto the lower Landau level, the vortex distribution
extends to infinity. This suggests that in this case, there are many vortices in the
low density region. It is then very natural to wonder whether vortices first appear
in the bulk or at infinity. It is experimentally and numerically difficult to observe a
vortex, which is a zero, in a low density region. Mathematically this could not be
achieved through energy estimates or expansion since the contribution of a vortex in
a low density region is very small. In this paper, we introduce new ideas to answer
Pitaevskii’s question and prove that at low velocity, there are indeed no vortices in
the condensate, even in the low density region. Therefore, the first ground state with
vortices has its vortices in the bulk.
Since a condensate is a trapped object, the geometry of the trap plays a role. An im-
portant special case is a radial harmonic trapping potential V (r) = r2 . The space can
then be split into two regions, a region of the form D = {λ0 > V (r)} (for a suitable
constant λ0 ), where the wave function is significant and the condensate is mainly
located, and a region R2 \D where the modulus of the wave function is exponentially
small [4]. In this latter region, it is very difficult to determine mathematically the
contribution of a vortex to the energy. Ignat and Millot [100, 101] have determined
the critical rotational velocity Ωc for the nucleation of the first vortex inside D. This
theorem does not describe the behaviour in R2 \ D. A natural question is whether
for Ω < Ωc , the minimizer of the energy has zeroes in this region, whether there is
a smaller critical velocity than Ωc where the minimizer is unique and vortex free.
At very high velocity, it has been proved in [7] that vortices exist up to infinity in a
reduced model so it seems reasonable that at smaller velocity, vortices may exist in
the exponentially small region, far away from the bulk and could arrange themselves
on disks or arrays close to infinity. In fact, we prove that this is not the case before
Ωc , namely that the minimizer is unique and does not vanish. It means that for a
large range of rotational velocities Ω, the minimizer exactly equals the ground state
of a condensate at rest.
We consider here a two-dimensional

R setting and define the energy for the complex-
2
valued wave function u, such that R2 |u| = 1, as
Z
1 2 1 4 1 2 ⊥
Eε (u) = |∇u| + 2 |u| + 2 V (x)|u| − Ωx · (iu, ∇u) dx, (2.1.1)
R2 2 4ε 2ε
where Ω is the angular velocity, x = (x1 , x2 ), x⊥ = (−x2 , x1 ), ε > 0 is a small

parameter, V (x) is the trapping potential and (iu, ∇u) = iu∇u∗ − iu∗ ∇u. The
class of potentials includes the model case V = x21 + x22 . Then, the critical angular
velocity for nucleation of vortices is of order | log ε| (see [100]). An upper bound on
the rotational velocity is given by O < 1/ε when the confinement breaks down. The
2.1. INTRODUCTION 51
condensate is mostly concentrated in the region
D := {x ∈ R2 : V < λ0 } (2.1.2)
where λ0 is chosen so that

Z
(λ0 − V (x))+ dx = 1. (2.1.3)
R2
We refer to [4] for more details on how this is derived from the physical experiments.
In recent experiments in which a laser beam is superimposed upon the magnetic

trap holding the atoms, the trapping potential V (x) is of a different type [154, 161,
167, 13] :
2
V (r) = r2 + V0 e−r /w0 . (2.1.4)
When the gaussian is expanded around the origin, this leads to a harmonic plus
quartic potential [161]
k
V (r) = (1 − b)r2 + r4 . (2.1.5)
4
If b is small (b < 1 + (3k 2 /4)1/3 ), the domain D given by (2.1.2) is a disc, while if
b > 1 + (3k 2 /4)1/3 , it is an annulus. According to the values of V0 and w0 in the case
of (2.1.4), the domain D can also be a disk or an annulus.
In this paper, we consider potentials V including r2 and of the type (2.1.4) or (2.1.5)
when the bulk D is a disk. In the case where D is a disk, the potential V is not
necessarily required to be increasing.
2.1.1 Assumptions
Throughout this paper, we make the following assumptions about the potential V .
First,
V is nonnegative and radial, V ∈ C 1 , (2.1.6)
and
1 p
there exists c0 > 0, p ≥ 2 such that r ≤ V (r) ≤ c0 rp if r ≥ c0 . (2.1.7)
c0
This assumption is easily seen to imply that Eε is bounded below for |Ω| 1ε and
that the angular momentum term x⊥ ·(iu, ∇u) is integrable as long as u has finite
energy. We will also use (2.1.7) to obtain decay estimates that justify for example
the integration by parts leading to a decoupling of the energy. We fix λ0 ∈ R such
that (2.1.2)-(2.1.3) hold. Such a λ0 exists due to the growth of V . We further assume
that the bulk D is a disk and not an annulus, that is V is such that
D = BR (0) for some R > 0 (2.1.8)

and that there exist δ0 > 0 and a C 1 function R : (−2δ0 , 2δ0 ) → R also denoted
Rδ = R(δ), such that
R0 = R , {x : V (x) < λ0 + δ} = BRδ (0)

(2.1.9)
1
and 0 < C
≤ dR/dλ ≤ C for some constant C
where Br (y) denotes the open ball of radius r about y. This implies that λ0 − V
is bounded away from 0 in the interior of D ; in physical terms, this assumption
rules out the case of annular bulks and “giant vortices" at low angular velocities. We
remark that the assumption above implies that if |x| ∈ (R−δ , Rδ ) and 0 ≤ δ ≤ δ0
then dist (x, ∂D) = O(δ).
We point out that assumptions (2.1.7) and (2.1.9) imply that
there exists c1 > 0 such that V (r) − λ0 ≥ c1 (r2 − R2 ) for all r ≥ R. (2.1.10)
Our assumptions include indeed potentials like r2 or (2.1.5) for a disk case, and do
not require V to be increasing.
2.1.2 Main result

Our main result is
Theorem 2.1. Assume that uε minimizes Eε (·) with rotation Ω, and let ηε denote
the minimizer of Eε (·) for Ω = 0. There exists ε0 , ω0 , ω1 > 0 such that if 0 < ε < ε0
and Ω ≤ ω0 | log ε| − ω1 log | log ε| then uε = eiα ηε in R2 for some constant α.
In the pure quadratic case V = r2 , Ignat and Millot [100, 101] have shown that
the bulk of the condensate (that is any domain contained in D) is vortex-free for
|Ω| ≤ ω0 | log ε| − ω1 log | log ε|, for some ω1 > 0 and the same constant ω0 that
we find in Theorem 2.1 . They have no information on what happens in R2 \ D.
Our theorem proves that vortices do not lie in R2 \ D. They have also shown that
there exists δ > 0 such that the ground state has at least one vortex in the bulk if
Ω ≥ ω0 | log ε| + δ log | log ε|. In this sense, our estimate |Ω| ≤ ω0 | log ε| − ω1 log | log ε|
captures the sharp leading-order term, and the correct scaling of the next-order term,
of the critical velocity for vortex formation. We point out that our arguments also
deal with more general potentials. The arguments used in [100] to prove the existence
of interior vortices for rotations greater than ω0 | log ε| + δ log | log ε| should extend
with few changes to the more general potentials considered here, using results about
auxiliary functions that we establish in Section 3 in place of parallel results from
[100]. Thus the constant ω0 should also be sharp for these more general potentials.
We split the proof into two independent results. The first main result of this paper
asserts roughly speaking that symmetry breaking occurs first in the interior of D :
if Ω is small enough that there are no vortices in D, then there are no vortices
anywhere, and in fact the rotation has absolutely no effect on the ground state.
Theorem 2.2. Assume that uε minimizes Eε (·) with rotation Ω, and let ηε denote
the minimizer of Eε (·) for Ω = 0. Assume also that Ω ≤ C| log ε| for some C.
There exists ε0 > 0 such that if 0 < ε < ε0 and Ω is subcritical in the sense that
1
|uε | ≥ ηε in D1 := {x ∈ D : dist(x, ∂D) ≥ | log ε|−3/2 } (2.1.11)
2
then uε = eiα ηε in R2 for some constant α.
Our second main theorem gives an estimate for the critical value of Ω. The statement
of the theorem refers to an auxiliary function f0 : let
a(x) = λ0 − V (x),
√ ∞
Z
0 if r ≥ R
η0 := a+ , ξ0 (r) = sη02 (s) ds, f0 (r) =
r ξ0 (r)/η02 (r) if r ≤ R;
(2.1.12)
Theorem 2.3. Let ω0 = 2kf10 k∞ . There exists ω1 > 0 and ε1 > 0 such that if
|Ω| ≤ ω0 | log ε| − ω1 log | log ε| and 0 < ε < ε1 , then Ω is subcritical in the sense of
(2.1.11), and the conclusion of Theorem 2.2 thus holds.
In our proof of Theorem 2.3, as in estimates of the critical rotation in works such
as [100] and [5], a main point is to obtain sharp energy lower bounds. In all earlier
works that we know of, this is done using the vortex ball construction originally
introduced by [103] and [155]. In our proof of Theorem 2.3, we avoid any explicit 4
mention of vortex balls by instead appealing to a result from [104], stated here as
Lemma 2.8. This makes our argument considerably shorter than those in [5, 100]
and other references.
We point out that the results of [100, 101] do not directly imply that Theorem 2.3
holds in the case V = r2 , although it is possible that this conclusion can be extracted
with relatively little effort from arguments in these references.
2.1.3 Main ideas of the proof

The energy minimizers with O = 0 provide real solutions to the Euler-Lagrange
equations when O = 0, Eε (η) = Gε (η), where
Z
1 2 1 4 1 2
Gε (η) = |∇η| + 2 |η| + 2 V (x)|η| dx. (2.1.13)
R2 2 4ε 2ε
Our main goal consists in proving that up to the critical velocity of nucleation of
bulk vortices, the minimizer of Eε with velocity Ω is in fact equal to ηε .
The minimizer ηε of Gε under the L2 constraint of norm 1, is (up to a complex

multiplier of modulus one) the unique positive solution of
1 1
− ∆ηε + 2
ηε (V (x) + ηε2 ) = 2 λε ηε (2.1.14)
ε ε
4. However, the proof of Lemma 2.8, see Lemma 8 in [104], ultimately relies on a vortex ball
construction appearing in [105].
where ε12 λε is the Lagrange multiplier, which is also necessarily unique. Moreover,
λε → λ0 , and ηε2 converges to a+ in L2 (D) and uniformly on any compact set of D.
We will need some estimates on the decay of ηε at infinity that we prove in section 2.
By a remarkable identity (see Lassoued & Mironescu [114]), for any u, the energy
Eε for any Ω splits into two parts, the energy Gε (ηε ) of the density profile and a
reduced energy of the complex phase v = u/ηε :
Eε (u) = Gε (ηε ) + Fε (v) (2.1.15)
ηε2 η4
Z
where Fε (v) = |∇v|2 + ε2 (|v|2 − 1)2 − ηε2 Ωx⊥ · (iv, ∇v) dx. (2.1.16)
R2 2 4ε
In particular the potential V (x) only appears in Gε . We will recall the proof of
(2.1.15), as well as that of (2.1.18) below, in Section 2.3. This kind of splitting of
the energy is by now standard in the rigorous analyses of functionals such as Eε .
Next, define Z ∞
ξε (r) = sηε2 (s) ds, (2.1.17)
r
so that ∇⊥ ξε = x⊥ ηε2 . An integration by parts yields

Z 2
ηε4

ηε 2 4Oξε 2 2
Fε (v) = |∇v| − 2 Jv + 2 (|v| − 1) dx (2.1.18)
R2 2 ηε 4ε
where Jv = 21 ∇ × (iv, ∇v) = (ivx1 , vx2 ) is the Jacobian. We recall that the func-
tion fε := ξε /ηε2 appearing in Fε is important since it is well known that vortices
in the interior of D first appear near where this function attains a local maximum
[4, 5, 100, 101] ; its importance is also clear from (2.1.18), since it controls the relative
strength of the positive and negative contributions to Fε . The proofs of Theorems
2.2 and 2.3 rest on new bounds for fε in R2 \ D and near ∂D, which in turn rely on
decay estimates for ηε . In particular, we show in Lemma 2.7 that fε ≤ Cε2/3 in R2 \D.
R
The other part of the proof consists essentially of bounds of 2Ω ηε2 fε Jv by the
positive terms in Fε . Away from the bulk, we use our estimates of fε to find that
2Ωfε Jv is bounded pointwise by 12 |∇v|2 . In the bulk, where ηε2 is not too small, we
have
1 2 η4 1 1
ηε |∇v|2 + ε2 (|v|2 − 1)2 ≥ ηε2 [ |∇v|2 + 2 (|v|2 − 1)2 ]
2 4ε 2 4ε̃
for some ε̃ such that | log ε̃| = | log ε|(1 + o(1)). We obtain the desired bounds by
combining this with a weighted Jacobian estimate mentioned above, Lemma 2.8,
which directly implies that
Z
2kfε k∞
Z
2 1 1
2Ω χηε fε Jv ≤ Ω χηε2 [ |∇v|2 + 2 (|v|2 − 1)2 ] + small error terms
| log ε̃| 2 4ε̃
where χ is a cutoff function supported
in the bulk. Note that the leading-order cri-
2kfε k∞
tical rotation ω0 is such that Ω log ε̃| ≈ Ω/ω0 | log ε|. The proof of Theorem 2.3 is
completed by assembling these ingredients and controlling error terms. The proof of
Theorem 2.2 relies on an additional ingredient, which is that if |v| ≥ 21 in an open set
2.2. PROPERTIES OF AUXILIARY FUNCTIONS 55
v
U , then Jv is extremely close in U to J( |v| ) = 0. Theorem 2.1 follows immediately
from combining Theorems 2.2 and 2.3.
An interesting open problem is to see to what extent this analysis continues to hold
if the assumption of radial symmetry is dropped. In our arguments, this symmetry
is used heavily in our analysis of the behavior of fε away from the bulk, and near
the boundary of the bulk.
We briefly remark on the assumption (2.1.7) of quadratic growth. Our proofs show
that the absence of vortices in the low density region is a consequence of the fact
that the auxiliary function fε = ξε /ηε2 is very small in R2 \ D. The proof of this
fact (see see Lemma 2.7) can be modified to show that if for example (2.1.7) holds
with p < 2, then fε (r) ≥ Cεr1−p/2 → ∞ as r → ∞. However, in this situation Eε is
unbounded below for any Ω 6= 0. This reflects the fact that a subquadratic trapping
potential is not strong enough to contain a rotated condensate.
2.2 Properties of auxiliary functions

In this section we study the real-valued minimizer ηε and the auxiliary functions ξε
and fε = ξε /ηε2 defined as
Z ∞
ξε (r) = sηε2 (s) ds, fε (r) = ξε (r)/ηε2 (r). (2.2.1)
r
Theorem 2.4. Assume that V satisfies (2.1.6), (2.1.9). Then for every ε > 0, there
exists a unique positive minimizer ηε of Gε in
Z Z
1 2 2
H := {u ∈ H (R ) : |u| V (x) < ∞, |u|2 = 1}.
R2 R2
Every minimizer of Gε in H has the form eiα ηε , for α constant. Moreover, ηε is a

radial smooth positive function and satisfies (2.1.14) with
|λε − λ0 | ≤ Cε| log ε|1/2 (2.2.2)
where λ0 is defined by (2.1.3). Finally, recall the notations Rδ from (2.1.9) and
a = λ0 − V , the following estimates are satisfied :
−1/3 (R−r)
ηε (r) ≤ C ε1/6 e c ε in R2 \ D (2.2.3)
√ √
|ηε − a+ | ≤ Cε1/3 a+ in BR−ε1/3 (2.2.4)
k∇ηε kL∞ (R2 ) ≤ Cε−1 . (2.2.5)
ηε0 (r) ≤ 0 for all r ∈ (R−δ0 , Rδ0 ) (2.2.6)
C p
|ηε0 (r)| ≤ ηε (r) V (r) for all sufficiently large r (2.2.7)
ε
if ε < ε0 .
Certain parts of the proof follow quite closely arguments given in [5] and in the
pure quadratic case in [100]. Note that some arguments in [100] rely strongly on the
special shape of the potential and cannot be generalized to other functions. Since V
is not necessarily increasing, we have property (2.2.6) only in the neighborhood of
∂D.
Démonstration. Step 1 : existence
R of minimizers : This follows from standard ar-
guments once we notice that R2 |un |2 V dx is uniformly bounded for any sequence
(un ) minimizing Gε , and the set of functions in H satisfying such a uniform bound
is precompact with respect to weak convergence in H 1 (R2 ). This last fact is proved
by straightforward and well-known arguments, such as are explained in theR proof
in [100], Lemma 2.1, for V quadratic, the point being that the bound on |u|2 V
prevents mass escaping to ∞. Standard theory then implies that any minimizer is
smooth. If η is any minimizer, then |η| is as well, since G(|ζ|) ≤ G(ζ) for all ζ.
The strong maximum principle then implies that |η| (and hence η) never vanishes,
and since G(η) ≤ G(|η|), it is easy to see that η/|η| = eiα for some constant α. We
henceforth let ηε denote a fixed positive minimizer.
Step 2 : uniqueness of ηε . Multiplying (2.1.14) by ηε and integrating by parts we

find that µε is positive. Suppose that there are two couples (η0 , µ0 ) and (η1 , µ1 )
satisfying (2.1.14) such that kη0 kL2 = 1 = kη1 kL2 and µ0 > µ1 , and define w = ηη10 .
This function verify
Z Z Z
2 2 2
η0 (w − 1) dx = 2 (η1 − η0 η1 ) dx = 2 η02 w(w − 1) dx
R2 R2 R2
and
1 4
−∇ · (η02 ∇w) + η w(w2 − 1) = (µ1 − µ0 )η02 w.
ε2 0
Multiplying the second equality by (w − 1), integrating by parts and then using the
first equality we find
Z
2 2 1 4 2 1 2 2
η0 |∇(w − 1)| + 2 η0 w(w − 1) (w + 1) + (µ0 − µ1 )η0 (w − 1) dx = 0.
R2 ε 2
The integration by parts is justified in view of (2.2.3), (2.2.7), which apply to both
η0 and η1 , and the proofs of which do not rely on the uniqueness of the minimizer.
Hence w ≡ 1 and µ0 = µ1 .
From (2.1.6) is easy to see that the compose of ηε with any rotation has the same
energy, so it is also a minimizer of Gε . The unicity implies then that ηε is a radial
function.
Step 3 : estimate of λε − λ0 . We next note, following standard arguments, that Gε

can be rewritten
Z Z
1 2 1 2 + 2 1 − 2 1 1 + 2
Gε (η) = |∇η| + 2 (η − a ) + 2 a η dx + 2 λ0 − (a )
R2 2 4ε 2ε 2ε 2
if kηk2 = 1. Let G1ε (η) denote the first integral above. We claim that
G1ε (ηε ) ≤ C| log ε|.
Since ηε is a minimizer, to prove this it suffices to construct a competitor for which

G1ε is suitably small. To do this, define
gε (a+ )
s
if s ≤ ε2
gε (s) := √ε and η̃ ε := .
s if s ≥ ε2 , kgε (a+ )kL2
Note that
a+
Z Z Z Z
+
1= a ≥ gε2 (a+ ) = +
a − a+ {1 − } ≥ 1 − Cε2 .
a+ ≤ε2 ε2
Using this and explicit calculations such as those in [104], Lemma 12, the claim
is easily verified. We now multiply (2.1.14) by ηε , integrate by parts and rewrite,
recalling the L2 constraint, to find that
Z
1 1
2
(λε − λ0 ) = |∇ηε |2 + 2 (ηε2 + (V − λ0 ))ηε2 dx (2.2.8)
ε ε
Z
1
= |∇ηε |2 + 2 (ηε2 − a+ + a− )ηε2 dx
ε
Z
1
|∇ηε |2 + 2 a− ηε2 + (ηε2 − a+ )2 + (ηε2 − a+ )a+ dx

= (2.2.9)
ε
1 1p 1
≤ 4G1ε (ηε ) + 2 kηε2 − a+ kL2 ka+ kL2 ≤ C[G1ε (ηε ) + Gε (ηε )].
ε ε
Thus we have proved (2.2.2).
Step 4 : estimates of ηε .
We claim that
ηε2 ≤ max(λε − V ) =: A (2.2.10)

D
√
To see this define w = 1ε (ηε − A). We have that ηε ∈ L3loc , so after (2.1.14)
w, ∆w ∈ L1loc . Kato’s inequality gives ∆w+ ≥ sgn+ (w)∆w. Using (2.1.14) again we
find
sgn+ (w) sgn+ (w) √ √

∆w+ ≥ 3
η ε (ηε
2
− A) = 3
(εw + A)(ε 2 2
w + 2εw A) ≥ (w+ )3 in D0
ε ε
Hence we have −∆w+ + (w+ )3 ≤ 0 in D0 (R2 ) and w ∈ L3loc , so using Lemma 2 in
[45], w+ ≡ 0.
We remark that the properties of the potential V at the boundary (2.1.9) implies
that the maximum of λε − V is attained at an interior point x0 of D such that
dist (x0 , ∂D) > c δ0 .
The minimizer being a solution of (2.1.14) in L∞ , by elliptic regularity we derive

that it is a smooth function.
Proof of (2.2.3). We construct a supersolution of (2.1.14) of the form
 p

 λ0 − V (x) + 8δ if |x| ≤ R−δ


√

λ0 −-.V (x)

η̄(x) := √
6 δ
+ 3 δ if R−δ ≤ |x| ≤ Rδ




 |x|
γ e− σ if Rδ ≤ |x|

where 0 < < . δ0 is small parameter that will be determined later and γ, σ are chosen
such that η̄ ∈ C 1 (R2 ), i.e.,
√
8 δ Rδ /σ 16 δ
γ= e and σ =
3 |∇V (Rδ )|
1/3
. Cε , η̄ is a supersolution of (2.1.14)
A straightforward computation shows that for =
and we also have
−1/3

σ = O(ε1/3 ) and γ = O ε1/6 eε R .
Moreover, with this choice of δ, η̄ 2 > λε − V for every |x| ≤ R−δ , so using (2.2.10)
ηε2 (x0 ) ≤ A = λε − V (x0 ) < η̄(x0 ).
Because ηε and η̄ are going to zero at infinity, if the function ηε − η̄ is positive

somewhere in (r0 , ∞), for r0 := |x0 |, then it attains a positive maximum at r̃ ∈
(r0 , ∞), i.e. ηε0 (r̃) = η̄ 0 (r̃) and ηε00 (r̃) < η̄ 00 (r̃). Given the structure of (2.1.14) and
because η̄ is a supersolution and ηε a solution, if V (r̃) − λε ≥ 0 we would have
p ηε (r̃) ≤ η̄(r̃). In another hand, if V (r̃) − λε < 0 then we would have η̄(r̃) <
that
λε − V (r̃), which for ε small enough, contradicts the definition of η̄. Hence
ηε (r) ≤ η̄(r) in (r0 , ∞) .

Proof p
of (2.2.4). Using
p assumption (2.1.9), by exactly following [5], one finds that
+ 1/3
|ηε − aε | ≤ Cε a+
ε , for aε := λε − V = a + λε − λ0 . In view of (2.2.2), this
implies (2.2.4).
Proof of (2.2.5). For x ∈ R2 define η̃(y) = ηε (ε(y − x)) in B2L (x). This function
satisfies
∆η̃ = η̃ (V (ε(y − x)) + η̃ 2 − λε ) =: hε
After estimates (2.2.3) and (2.2.4) |hε | ≤ C, so using a Hölder estimate for the first
derivative of η̃ (see Theorem 8.32 in [80]) we have that k∇η̃kL∞ (BL (x)) ≤ C for a
constant C independent of x and hence the result.
Step 5 : proof of (2.2.6).
We denote L the elliptic operator obtained by linearizing equation (2.1.14)

1
L := −∆ + (V (x) + 3ηε2 − λε ),
ε2
and λj , j = 1, 2, ... , its eigenvalues in R2 .
Let µ be the first Dirichlet eigenvalue of L in the half space Ω = {x1 > 0} and ψ
the corresponding eigenfunction (which exists because of the compact embedding of
H in L2 ). Since V and ηε are radial, is clear that the odd extension of ψ to R2 is
a eigenfunction for L in R2 with corresponding eigenvalue µ = λj . Note that j ≥ 2
because the odd extension change sign in R2 .
We have that Lηε = 2ηε4 > 0 and ηε > 0. Using the maximun principle due to
Berestycki, Nirenberg and Varadhan [32], this implies that the first eigenvalue of L
is positive. We will prove that if (2.2.6) does not hold, then µ < 0, which contradicts
the fact that λ1 > 0. Assume that ηε0 (r) > 0 at some r ∈ (R−δ0 , Rδ0 ). Then there
exists α < r < β such that ηε0 (α) = ηε0 (β) = 0 and ηε0 > 0 in (α, β). If α ≤ R−2δ0 , then
ηε is increasing on (R−2δ0 , R−δ0 ), so that√ ηε (R−2δ0 ) ≤ ηε (R−δ0 ). This is impossible
for all sufficiently small ε, since ηε → a+ uniformly for r < R−ε1/3 , by (2.2.4),
and a+ (R−2δ0 ) > a+ (R−δ0 ). Thus α ≥ R−2δ0 . The same argument, but using (2.2.3)
instead of (2.2.4), shows that β ≤ R2δ0 .
Now let D := {x ∈ R2 : x1 > 0, α < |x| < β}. Then

∂ηε ∂ηε ∂ηε ∂V
> 0 in D , = 0 in ∂D and L =− ηε ≤ 0 in D .
∂x1 ∂x1 ∂x1 ∂x1
The last inequality come from the differentiation of (2.1.14) and hypothesis (2.1.9),
which implies that ∂V /∂R > 0 for r ∈ (R−2δ0 , R2δ0 ). Using the monotonicity of
Dirichlet eigenvalues with respect to the domain, this implies that µ < 0.
Step 6 : proof of (2.2.7). For any r ≥ R, define a function η̃ : (r, ∞) → R by

2α p+2 p+2
η̃(s) := ηε (r) exp − (s 2 − r 2 )
p+2
where c0 and p are the constants in (2.1.7). It follows from (2.2.2) and (2.1.7) that
if s ≥ r and r is sufficiently large, then V (s) − λε + η̃ 2 (s) ≤ V (s) ≤ c0 sp , so that if
r is sufficiently large, then
1 2 c0 p
2 c0 p p p
−1

−∆η̃ + 2 (V − λε + η̃ )η̃ ≤ −∆η̃(s) + 2 s η̃ = (−α + 2 )s + α( + 1)s 2 η̃.
ε ε ε 2
1/2
Choosing α = (2c0ε) , it follows that η̃ is a subsolution of (2.1.14) in (r, ∞) if r is
sufficiently large. For such r, noting that η̃(r) = ηε (r), we can argue as in the proof
of (2.2.3) to deduce that ηε − η̃ is nonnegative in (r, ∞).
Then since η̃(r) = ηε (r) and η̃(s) ≤ ηε (s) for s ≥ r, we again use (2.1.7) to conclude
that
(2c0 )1/2 p √ c0 p
ηε0 (r) ≥ η̃ 0 (r) = − r 2 ηε (r) ≥ − 2 V (r)ηε (r)
ε ε
√
c
for sufficiently large r. On the other hand, by choosing α = 2ε0 in the definition
of η̃, we obtain a decreasing supersolution (still denoted η̃) such that η̃(r) = ηε (r).
A similar application of the maximum principle shows that ηε is bounded above by
(the new) η̃ on (r, ∞), and in particular this implies that ηε0 (r) ≤ 0. These facts
combine to establish (2.2.7).
We next prove
Lemma 2.5. Assume that V satisfies (2.1.6) and (2.1.9) and the quadratic growth
condition (2.1.10). Let ηε be the positive minimizer found in Theorem 2.4. Let
fε (r) := ξε (r)/ηε2 (r), where ξε was defined in (2.1.17). Then there exists a constant
C independent of ε ∈ (0, ε1 ] such that

Cdist (x, ∂D) + Cε2/3 if x ∈ D
fε (|x|) ≤ (2.2.11)
Cε2/3 if not.
In addition, for all sufficiently small ε,
k∇ξε k∞ ≤ C (2.2.12)
and
kfε − f0 k∞ ≤ Cε1/3 . (2.2.13)
Démonstration. For every s ≥ r ≥ Rδ (where 0 < δ ≤ δ0 will be chosen later), we

define
2 −r 2 )/2 c1 (Rδ2 − R2 ) + (λε − λ0 )
η̃(s) = ηε (r)e−µδ (s and µ2δ = . (2.2.14)
Rδ2 ε2
Using (2.1.10), where the constant c1 is defined, and arguing as in the proof of
(2.2.7), we find that η̃ − ηε is nonnegative in (r, ∞).
We use the previous estimate and the definition of ξε to compute
Z ∞ Z ∞
1 2 2 1
fε (r) = 2 2
sηε (s) ds ≤ e−µδ (s −r ) s ds = for r ≥ Rδ .
ηε (r) r r 2µδ
0
The definition of fε implies that fε0 (r) = −r −2fε (r) ηηεε (r)
(r)
, and from the monotonicity
0
(2.2.6) of ηε , we infer that fε (r) ≥ −r in (R−δ0 , Rδ0 ). Thus for any R−δ0 ≤ r ≤ Rδ ,
Rδ2 − r2 1
fε (r) ≤ + .
2 2µδ
We now fix δ = ε2/3 , and we conclude from (2.1.9) and (2.2.2) that (2.2.11) holds as
long as r ≥ R−δ0 .
For 0 ≤ r ≤ R−δ0 , we write
R−δ0
ηε2 (R−δ0 )
Z
1
fε (r) = sηε2 (s) ds + f (R−δ0 )
ηε (r)2 r ηε2 (r)
From (2.2.4) and (2.1.9), we see that if 0 ≤ r ≤ s ≤ R−ε1/3 , then
ηε2 (s) (1 + Cε1/3 )2 a+ (s)

≤ ≤C for sufficiently small ε, (2.2.15)
ηε2 (r) (1 − Cε1/3 )2 a+ (r)
and by using the and the fact that fε (R−δ0 ) ≤ Cε2/3 + Cδ0 , one easily deduces that
(2.2.11) holds for r ∈ [0, R−δ0 ).
Next, the definition of ξε implies that |∇ξε (x)| = |x|ηε2 (x), so that (2.2.12) follows
from (2.2.10) and (2.2.3).
For r ≥ R−ε1/3 , we see from (2.2.11) that |fε (r) − f0 (r)| ≤ Cε1/3 + |f0 (r)|. This is
trivially bounded by Cε1/3 if r ≥ R. If R−ε−1/3 ≤ r ≤ R then (2.1.9) implies that
c(R − r) ≤ a(r) ≤ C(R − r), and thus
Z R
C
|f0 (r)| = f0 (r) ≤ s(R − s)ds ≤ C(R − r) ≤ Cε1/3 .
r−R r
For 0 ≤ r ≤ R−ε1/3 we write
Z R 1/3 Z R 1/3
1 −ε
2 1 −ε
nnfε (r) − f0 (r) = sηε (s) ds − sa(s) ds (2.2.16)
ηε2 (r) r a(r) r
η 2 (R 1/3 ) a(R−ε1/3 )
+ ε 2 −ε fε (R−ε1/3 ) − f0 (R−ε1/3 ) (2.2.17)
ηε (r) a(r)
= I + II − III (2.2.18)
nn Using (2.2.15) and our earlier estimates of fε , f0 for r ≥ R−ε1/3 , we see that
|II| ≤ Cfε (R−ε1/3 ) ≤ Cε1/3 and |III| ≤ Cf0 (R−ε1/3 ) ≤ Cε1/3 .
We further decompose the remaining term as
Z R 1/3 Z R 1/3
1 1 −ε
2 1 −ε
I= 2
− sηε (s) ds + s(ηε2 (s) − a(s)) ds.
ηε (r) a(r) r a(r) r
Using (2.2.4), it follows that

Z R 1/3 2 Z R 1/3
1/3
−ε ηε (s) 1/3
−ε a(s)
|I| ≤ Cε s 2 ds + Cε s ds.
r ηε (r) r a(r)
ηε2 (s)
Due to (2.2.6), ηε2 (r)
≤ 1 if R−δ0 ≤ r ≤ s ≤ R−ε1/3 . And if 0 ≤ r ≤ R−δ0 then
2
≥ C and so ηηε2 (r)
ηε2 (r) −1 (s)
≤ C. Thus the first integral is bounded by Cε1/3 . The
ε
second integral is similarly estimated, using (2.1.9) in place of (2.2.6).
Remark 2.6. In the case of a potential V for which (2.1.8) fails, so that for example
D has the form BR \ BR0 , one expects that instead of being small, fε is large, namely,
fε ≥ cec/ε in the interior of BR0 . This is related to the formation at very low rotations
of a giant vortex in the interior of BR0 . The arguments used to prove Lemma 2.7
show in this situation that if V grows quadratically in the complement of BR , as in
(2.1.10), then fε is very small in R2 \ BR . This suggests that at low rotations there
should be no vortices in R2 \ BR , but this cannot be deduced from the arguments we
use to prove Theorems 2.2 and 2.3.
The last lemma in this section examines the case when V has subquadratic growth
and fε is also large so that in principle vortices could exist in the low density region.
Lemma 2.7. Assume that V satisfies (2.1.6), (2.1.9) and
there exists c2 > 0 and p < 2 such that V (r) ≤ c2 (rp + 1) for all r ≥ R. (2.2.19)
Then fε (x) → +∞ as |x| → ∞.
Note that with these assumptions on V , there is a sequence of functions ζα in H such

that inf α Gε (ζα ) = −∞. Physically this happens because the centrifugal force due
to rotation is bigger than the subquadratic trapping potential. This indicates that,
although one can prove that in this situation, fε → ∞ as r → ∞ (compare Lemma
2.7), this is not expected to give any information about the physical behaviour of
condensates.
Démonstration. Let q > 2. For every r ≥ max{1, R}, we claim that
q −r q )/q
ηε (s) ≥ ηε (r)e−νε,r (s (2.2.20)
q c
for all s ≥ r. Where νε,r is the positive root of the polynomial ν 2 − rq
ν − ε2 r2q−2−p
,
which for ε small satisfy
νε,r < C ε−1 r−β

with β = q − 1 − p/2. Indeed, the right hand side of (2.2.20) is a subsolution in
(r, ∞) of (2.1.14) while ηε is a solution. Boths functions are going to zero at infi-
nity and they are equal at s = r, so the result come arguing as in the proof of (2.2.3).
We use the previous estimates and the definition of ξε to compute

Z ∞ Z ∞
ξε (r) 1 q q r2−q
fε (r) = 2 = 2 2
sηε (s) ds ≥ e−νr (s −r ) s ds ≥ > C ε r1−p/2 .
ηε (r) ηε (r) r r ν r
and hence the result.
2.3 Splitting the energy

In this section we recall the proofs of (2.1.15) and (2.1.18). For U ⊂ R2 , we will
write Eε (w; U ) etc to denote the integrals over U of the energy density appearing in
the definition of Eε (u) = Eε (u; R2 ) , and similarly Gε (·, U ), Fε (·, U ).
Note that v = u/ηε is well defined since ηε > 0. Since ηε satisfies (2.1.14), we multiply
it by ηε (1 − |v|2 ) and integrate over a ball Br to find that
Z Z
2 1 2 1 2 2 1 2 λε
(|v| − 1)(− ∆ηε + 2 ηε (V (x) + ηε ) + |∇ηε | ) = 2 (|u|2 − ηε2 ).
Br 4 2ε 2 ε Br
Note that the Lagrange multiplier term tends to 0 as r → ∞, since both the L2
norms of u and ηε are 1. Moreover,
Z
1 1
Eε (vηε ; Br ) = Gε (ηε ; Br ) + Fε (v; Br ) + |∇ηε |2 (|v|2 − 1) + ηε ∇ηε · ∇|v|2
Br 2 2
1 1 1 1 1
− 2 ηε4 (1 − |v|2 )2 + 2 η 4 |v|4 + 2 V (x)η 2 |v|2 − 2 η 4 − 2 V (x)η 2 .
4ε 4ε 2ε 4ε 2ε
We integrate by parts to obtain
Z Z Z
1 2 1 2 2 1 2
ηε ∇ηε · ∇|v| = − |v| ∆ηε + |v| ηε ν · ∇η
Br 2 Br 4 ∂Br 2
2.4. PROOFS OF THEOREMS 2.2 AND 2.3 63
We use (2.2.7) to estimate

1 2 2√
Z Z Z
1 2 C C −1/2
| |v| ηε ν · ∇η| ≤ η |v| V = V (r) V |u|2 .
∂Br 2 ε ∂Br 2 ε ∂Br
R
Since R2 V |u|2 < ∞, we can easily find a sequence rk → ∞ such that the above
integral tends to 0. Combining the above and letting rk → ∞ along this sequence, we
obtain (2.1.15). The only property of V that the above argument used (implicitly)
was (2.1.7), which will be used in the proof of (2.2.7). The integration by parts that
leads to (2.1.18)
R is justified in a similar fashion. One must estimate boundary terms
of the form ∂Br ξν · (iv, ∇v). To do this we note that
ξν · (iv, ∇v) = fε (r)ηε2 (iv, ∇v) = fε (r)(iu, ∇u) ≤ kfε k∞ (|u|2 + |∇u|2 ).
We prove in (2.2.11) that fε is bounded as long as V satisfies (2.1.10) (in fact we
show that fε ≤ Cε2/3 for large r) and since u ∈ H 1 (R2 ), we can again find a sequence
rk → ∞ such that the boundary terms vanish. Note also that the fact that fε ∈ L∞ ,
or equivalently that |ξε | ≤ Cηε2 , implies that the term ξε Jv appearing in (2.1.18) is
integrable on R2 for v = u/ηε , whenever u has finite energy.
2.4 Proofs of Theorems 2.2 and 2.3

In this section we use the estimates we have already established to complete the
proofs of our main theorems.
Proof of Theorem 2.2. We assume that uε minimizes Eε and that Ω ≤ C| log ε| is
such that (2.1.11) holds.
Let χ be a smooth function such that χ ≡ 1 in {x ∈ D : dist(x, ∂D) ≥ 2| log ε|−3/2 },
and with support in D1 . We also assume that k∇χk∞ ≤ 2| log ε|3/2 .
Let v = uε /ηε , so that Eε (u) = Gε (ηε ) + Fε (v) = Eε (ηε ) + Fε (v). Thus Fε (v) ≤ 0.
We write
Fε (v) = A1 − A2 + B
where
ηε2 ηε4
Z Z
A1 = χ 2
|∇v| + 2 (|v|2 − 1)2 dx, A2 = 2O χξε Jvdx
R2 2 4ε R2
and
η2 η4
Z
(1 − χ) ε |∇v|2 − 4Ofε Jv + ε2 (|v|2 − 1)2

B= dx,
R2 2 4ε
It follows directly from our estimates on fε that 0 < fε ≤ C(ε2/3 + | log ε|−3/2 ) in the
support of 1 − χ, for small enough ε. Since Ω ≤ C| log ε|, it follows that Ofε ≤ 41 for
all sufficiently small ε and (recalling that |Jv| ≤ 21 |∇v|2 ) we deduce that
1
|∇v|2 − 4Ofε Jv ≥ |∇v|2
2
in the support of 1 − χ. It follows immediately that
2
ηε4
Z
ηε 2 2 2
B≥ (1 − χ) |∇v| + 2 (|v| − 1) dx ≥ 0 (2.4.1)
R2 4 4ε
and hence that B = 0 if and only if v is a constant of modulus 1 in the support of

1 − χ.
Since Fε (v) ≤ 0, it is clear that A1 + B ≤ A2 .
Next, define ε̃ = ε/(inf D1 ηε ), so that (in view of (2.2.4) and the definition of D1 )
1 ηε2
ε̃ ≤ Cε| log ε|3/4 , ≤ in D1 .
ε̃2 ε2
Then (2.4.1) and (2.2.4) imply that,
Z
1 1
|∇v|2 + 2 (|v|2 − 1)2 ≤ (inf ηε )−2 (A1 + 2B) ≤ C| log ε|3/2 A2 . (2.4.2)
D1 2 4ε̃ D1
v
To continue, let w = |v| = w1 + iw2 . From (2.1.11) we see that |v| ≥ 21 in D1 , and
hence it is clear that w ∈ H 1 (D1 ), and |w|2 ≡ 1. It follows that Jw = 0 ; we will
recall a standard proof of this fact in a moment. Thus
Z Z
A2 = 2Ω χξε (Jv − Jw) dx = 2Ω ∇⊥ (χξε ) · [(iv, ∇v) − (iw, ∇w)] dx.
D1 D1
If we write v = ρeiφ in D1 , then a calculation shows that
(iv, ∇v) = ρ2 ∇φ, (iw, ∇w) = ∇φ.
From the latter fact we see that Jw = 21 ∇ × (iw, ∇w) = 0, as we asserted above.
Also, from this and the fact that ρ ≥ 21 in D1 we estimate
|ρ2 − 1|
|(iv, ∇v) − (iw, ∇w)| = |ρ∇φ| ≤ 2| |v|2 − 1| |∇v|.
ρ
Using (2.4.2) , we deduce that

Z
ε̃ 2 1 2 2
nnA2 ≤ 2Ωk∇(χξε )k∞ |∇v| + (|v| − 1) dx (2.4.3)
D1 2 2ε̃
≤ CΩk∇(χξε )k∞ ε| log ε|9/4 A2 . (2.4.4)
nn
One checks easily from the definitions and from (2.2.12) that
k∇(χξε )k∞ ≤ k∇χk∞ kξε k∞ + k∇ξε k∞ ≤ C| log ε|3/2 (2.4.5)
so we conclude that A2 ≤ Cε| log ε|15/4 A2 ≤ 21 A2 for all sufficiently small ε. We know
from (2.4.2) that A2 ≥ 0, and it follows that A2 = 0, and hence (again appealing to
(2.4.2)) that A1 = B = 0. Thus k∇vkL2 = k1 − |v|2 kL2 = 0, and so v is a constant
of modulus 1 as required.
The proof of Theorem 2.3 will use the following result, which is Lemma 8 in [104].
2.4. PROOFS OF THEOREMS 2.2 AND 2.3 65
Lemma 2.8. There exists a universal constant C > 0 such that for any κ ∈ (1, 2),
open set U ⊂ R2 and u ∈ H 1 (U ; R2 ), and ε ∈ (0, 1),
Z
φJu ≤ κ |φ| eε (u)
R
U
| log ε|
Z
(κ−1)/50
+Cε (1 + kφkW 1,∞ ) kφk∞ + 1 + (|φ| + 1)eε (u) dx
(2.4.6)
supp φ
for all φ ∈ Cc0,1 (U ). Here eε (u) = 21 |∇u|2 + 1

4ε2
(|u|2 − 1)2 .
The lemma as stated in [104] does not explicitly specify the exponent (κ − 1)/50
appearing on the right-hand side of (2.4.6). By inspection of the proof, however, one
sees that this exponent can be taken to have the form 21 α, where α = (κ − 1)/12κ
as in Theorem 2.1 of [105].
Proof of Theorem 2.3. We continue to use notation from the proof of Theorem 2.2,
such as A1 , A2 , B, ε̃, and so on. We first invoke the lemma, with ε̃ in place of ε and
χ ξε in place of φ, and with κ > 1 to be chosen. This yields
Z
eε̃ (v)
|A2 | ≤ 2Ωκ χ ξε dx + E,
R2 | log ε̃|
where E denotes the error terms in (2.4.6). We note that for all sufficiently small
ε > 0, the error term satisfies the bound E ≤ Cεβ (1 + |A2 |), for β = (κ − 1)/100,
for all sufficiently small ε. This is a consequence of (2.4.2) and the estimates
kχ ξε kW 1,∞ ≤ C| log ε|3/2 , kχ ξε kL∞ ≤ C.
These in turn follow from (2.4.5) together with (2.2.12). Now the choice of ε̃ implies
ηε2
that eε̃ (v) ≤ 21 |∇v|2 + 4ε 2 2 2
2 (|v| − 1) in D1 , and recalling that ξε = fε ηε , we obtain
kfε k∞ ηε2 ηε4

Z
β
(1 − Cε )|A2 | ≤ 2Ωκ χ( |∇v| + 2 (|v|2 − 1)2 ) + Cεβ
2
| log ε̃| 2 4ε
kfε k∞
= 2Ωκ A1 + Cεβ .
| log ε̃|
We know from (2.2.13) that kfε k∞ ≤ (1 + Cε1/3 )kf0 k∞ ≤ (1 + Cεβ )kf0 k∞ , and from
the choice of ε̃, for any K > 0 there exists ε0 > 0 such that | log ε̃| ≥ (| log ε| −
log | log ε|)(1 + Kεβ ) if 0 < ε < ε0 . Thus

2kf k∞
|A2 | ≤ Ω κA1 + Cεβ
| log ε| − log | log ε|
for all sufficiently small ε. Assume that Ω ≤ 2kf1k∞ (| log ε| − (c1 + 1) log | log ε|), for
c1 to be chosen below. Then

log | log ε| β log | log ε|
|A2 | ≤ 1 − c1 κA1 + Cε ≤ 1 − c1 κA1 + Cεβ .
| log ε| − log | log ε| | log ε|
(2.4.7)
| log ε| c1 log | log ε|

We now take κ := 1 + c1 log| log ε|
, so that β = (κ − 1)/100 = 100 | log ε|
. Recalling
that A1 + B ≤ A2 and that B ≥ 0, clearly A1 ≤ A2 , so we deduce that
log | log ε| 2
c21 ( ) A1 ≤ Cεβ = C| log ε|−c1 /100 .
| log ε|
If c1 = 400 then we conclude that A1 ≤ C| log ε|−2 . Then (2.4.7) implies that
A2 ≤ C| log ε|−2 , and it follows that B ≤ C| log ε|−2 . In view of (2.4.2), this implies
that Z
1
|∇v|2 + 2 (|v|2 − 1)2 ≤ C| log ε|−2 . (2.4.8)
D1 4ε
C
The estimate k∇vk∞ ≤ ε
(see (2.2.5)) and (2.4.8) are easily seen to imply that
|v| ≥ 1 − C| log ε|−1 in D1 (2.4.9)
for all sufficiently small ε. Thus Ω is subcritical for small enough ε.

Chapitre 3
Segregation and symmetry breaking

of strongly coupled two-component
Bose-Einstein condensates in a
harmonic trap
Jimena Royo-Letelier 1
Article accepté le 22 octobre 2012 dans

Calculus of Variations and Partial Differential Equations
Abstract
We study ground states of two-component condensates in a harmonic trap.
We prove that in the strongly coupled and weakly interacting regime, the
two components segregate while a symmetry breaking occurs. More preci-
sely, we show that when the intercomponent coupling strength is very large
and both intracomponent coupling strengths are small, each component is
close to the positive or the negative part of a second eigenfunction of the har-
monic oscillator in R2 . As a result, the supports of the components approach
complementary half-spaces, and they are not radially symmetric.
3.1 Introduction
A two-component Bose-Einstein condensate (BEC) is described in terms of two wave
functions u and v, respectively representing the first and the second component.
The energy of a trapped two-dimensional two-component BEC is given for g =
(g1 , g2 , g12 ) by
Z
1 n 1 o n 1 o
Eg (u, v) = |∇u| + V |u| + g1 |u| + |∇v| + V |v| + g2 |v| + g12 |u|2 |v|2 .
2 2 4 2 2 4
2 R2 2 2
1. Laboratoire de Mathématiques de Versailles, CNRS UMR 8100, 45 avenue des États-Unis,
78035 Versailles Cédex, France.
67
68 CHAPITRE 3. SEGREGATION AND SYMMETRY BREAKING
Here g12 is the intercomponent coupling strength, and g1 (respectively g2 ) is the

intracomponent coupling strength of the first (respectively second) component. We
are interested in the properties of the ground state of Eg when g12 goes to infinity.
We assume that both components have repulsive internal interactions, so
g1 ≥ 0 and g2 ≥ 0 , (3.1.1)
and that the trapping potential is the harmonic function
V (x) = ω 2 |x|2 , ω > 0. (3.1.2)

We consider then (ug , vg ) as a minimizer of Eg over the class of finite energy pairs
with constrained mass
n o
1 2
X = (u, v) ; u, v ∈ HV (R ; C) , kukL2 (R2 ) = 1, kvkL2 (R2 ) = 1 ,
where
n Z o
HV1 (R2 ; C) 1 2
= u ∈ H (R ; C) ; V |u|2 < ∞ .
R2
The pair (ug , vg ) satisfies the system of coupled Gross-Pitaevskii equations

−∆u + V u + g1 |u|2 u + g12 |v|2 u = λu
(3.1.3)
−∆v + V v + g2 |v|2 v + g12 |u|2 v = µv
in R2 , where λ and µ are the Lagrange multipliers associated with the mass constraints.
The system (3.1.3) is a particular case of nonlinear elliptic systems with competition.
This denomination is due to the terms g12 |v|2 u and g12 |u|2 v, which depending on the
value of g12 , favor solutions that coexist or that spatially separate. The segregation
problem consists in studying the properties of the solutions when g12 is positive and
very large, since in this case, solutions tend to have disjoint supports.
The segregation of two-component BECs has been experimentally observed. The

achievement of two-component BECs constitutes an important research subject in
experimental physics (see for example [85], [127] or [146]). Two-component BECs
have been realized in several ways, using different types of particles : two different
isotopes of the same atom, isotopes of two different atoms, or a single isotope in
two different hyperfine states. The choice of the number of particles and the types
of isotopes determine the value of g1 and g2 . The value of g12 can be altered by
changing the hyperfine state of a portion of the particles, via applied magnetic or
optical fields. This allows large and tunable values for g12 ([35], [146], [163]). The-
refore, manipulating the values of the coupling strengths, it is possible to realize
segregated two-component BECs with different spatial configurations ([85], [146]).
We refer to [109] for the description of the physics.
In numerical analysis ([109], [110], [126]), several patterns have been observed for the
supports of two-component BECs. In [126], Mason and Aftalion perform a numeri-
cal analysis on the solutions of (3.1.3). They classify all the possible configurations
depending on g12 and g1 6= g2 . For the segregated case, they exhibit two main types
of patterns : symmetry preserving and symmetry breaking ones. In the first case,
one component is a disc centered in the minimum of V , and the other component is
an annulus surrounding the disc. In the second case, both components are close to
half balls.
In this paper, we address both segregation and symmetry breaking for two-component
BECs in all R2 . Our goal is to prove that when g12 goes to infinity, the supports of
ug and vg are disjoint ; and that when g1 and g2 go to zero, they break the symmetry
of the harmonic potential by approaching half-spaces. To our knowledge, the are no
results in the mathematical literature about symmetry breaking for two-component
BECs in the repulsive case.
Our method consists in two steps. First, we consider sequences g n = (g1n , g2n , g12
n
) such
n n n
that g12 goes to infinity and g1 (respectively g2 ) converges to some nonnegative limit
g1 (respectively g2 ). We prove that the associated sequence of minimizers (ugn , vgn )
converges to a limiting pair, which minimizes the addition of the Gross-Pitaevskii
energy of each component
Z Z
1 n 1 o 1 n 1 o
Eg1 ,g2 ,∞ (u, v) = |∇u| + V |u| + g1 |u|4 +
2 2
|∇v|2 + V |v|2 + g2 |v|4
2 R2 2 2 R2 2
(3.1.4)
over
n o
Y = (u, v) ∈ X ; u · v = 0 , (3.1.5)
the subset of X of fully segregated pairs. Then, we study the ground states of Eg1 ,g2 ,∞
when g1 and g2 are both equal to zero, which results in an eigenvalue problem that
leads to half-space geometry.
Our first result is,
Theorem A. Let (ugn , vgn )n∈N be a sequence of minimizers of Egn over X, with
n
g n = (g1n , g2n , g12 ) ∈ [0, c0 ]2 × R+ such that
lim g1n = g1 , lim g2n = g2 and n

lim g12 = ∞. (3.1.6)
n→∞ n→∞ n→∞
There exist a limiting pair (u∞ , v∞ ) ∈ Y such that, up to a subsequence,

(i) (ugn , vgn ) converges to (u∞ , v∞ ) weakly in HV1 (R2 ) × HV1 (R2 ) and in Cloc
0
(R2 ) ×
0 2
Cloc (R ).
(ii) (u∞ , v∞ ) minimizes Eg1 ,g2 ,∞ over Y and satisfies weakly the system

 −∆u + V u + g1 |u|2 u = λu in {|u∞ | > 0}
−∆v + V v + g2 |v|2 v = µv in {|v∞ | > 0} (3.1.7)
u·v = 0 in R2 .

Here λ and µ are respectively the limits of the Lagrange multipliers λn and µn ,
associated with (ugn , vgn ) by (3.1.3).
This is the analogue of the results of Wei and Weth in [165] for solutions of the
system (3.1.3) in bounded domains and without trapping potential. We obtain the
corresponding results by adapting the techniques in this papers to our setting. In
[165] (see also the work of Chang, Lin, Lin and Lin in [53]), the authors prove the
segregation in the strongly coupled case and the local uniform convergence to a limi-
ting pair solving the system (3.1.7). In [60], Conti, Terracini and Verzini study the
equivalent of the energy Eg1 ,g2 ,∞ defined in bounded domains and without trapping
potential. They prove the existence of minimizers and their Lipschitz regularity, and
give extremality conditions in the form of a system of subsolution of elliptic equa-
tions. We emphasize that we address the problem that actually corresponds to the
physical situation. In the previous mentioned works, no mathematical results are
presented about the symmetry breaking for solutions of system (3.1.3).
The main result of this article, and the improvement with respect to the previous
works, is to prove the symmetry breaking and to give an accurate description of the
behavior of the ground states, in the strongly coupled and weakly interaction case :
Theorem B. Let (ugn , vgn ) be a sequence of minimizers of Egn over X, with g n =

(g1n , g2n , g12
n
) such that g1n → 0, g2n → 0 and g12
n
→ ∞. Then, up to a subsequence, ugn
and vgn converge in Cloc (R ) respectively to e + wν+ and eiθ− wν− , where
0 2 iθ
2 ω|x|2
wν (x) = √ ω (x · ν) e− 2 , ν ∈ S1 (3.1.8)
π
is a second eigenfunction of the harmonic oscillator −∆ + ω 2 |x|2 in R2 , and θ+ , θ−
are real constants.
The limiting functions wν+ and wν− are not radial since they are supported in the
half-spaces {x · ν > 0} and {x · ν < 0}. Theorem B says that in compacts domains
of R2 , the supports of ug and vg approach these half-spaces when g1 → 0, g2 → 0
and g12 → ∞. The desired result about the symmetry breaking follows immediately :
Corollary 1.1. There are positive constants g0 and G, such that if (ug , vg ) is a
minimizer of Eg in X with
max{g1 , g2 } ≤ g0 and g12 ≥ G ,

then ug and vg are not radially symmetric.
Theorem B follows directly from the local uniform convergence stated in Theorem
A, together with an accurate description of the segregated minimizers in the non
interacting limit when g1 = g2 = 0, given by :
Theorem C. Let (u0 , v0 ) be a minimizer of E0,0,∞ over Y . Then, there is ν ∈ S 1

such that
(i) u0 is supported in {x · ν > 0} and v0 in {x · ν < 0}, and
(ii) u0 and v0 are the positive and the negative parts of wν , a second eigenfunction
of the harmonic oscillator −∆ + ω 2 |x|2 in R2 given by (3.1.8).
We prove Theorem C using the results of Ehrhard in [71] about the extremality
properties of half-spaces. Theorem C says that in the limit case when g1 = g2 = 0,
the nodal set of a segregated two-component BEC is a straight line. We emphasize
that in the case when g1 and g2 are small and positive, the exact configuration of
both components is still unknown, but we expect the nodal set to be an infinite
line with small curvature. We mention the work of Berestycki, Lin, Wei and Zhao in
[31], where they study the profile of the components near the interface using blow-up
techniques.
The study of the geometric nature of the support of the segregated minimizers of
E0,0,∞ is an example of an optimal partition problem, or of a free boundary problem
between several components.
The optimal partition problem consists in finding a partition of Ω ⊂ Rn by k disjoint

open sets ω1 , . . . , ωk , that minimizes a function of the ground state energies in each
component. This is an open problem in general, and only few results are known
about the exact configurations of thesePoptimal partitions. In [50], Caffarelli and
Lin consider the problem of minimizing ki=1 λ(ωi ), where λ(ω) is the first Dirichlet
eigenvalue of the Laplacian on ω. For Ω bounded, they show the existence of clas-
sical solutions as well as the regularity of the interfaces. In [90], Helffer, Hoffmann-
Ostenhof and Terracini also consider the problem of minimizing maxi=1,...,k µ(ωi ),
where µ(ω) is the first Dirichlet eigenvalue of some Schrödinger operator on ω. They
study the relation between nodal domains, spectral partitions and spectral proper-
ties in a bounded domain. An important question is whether a minimal partition is
nodal, that is, consisting in the nodal domains of an eigenfunction of the Dirichlet
realization of the Schrödinger operator in Ω. This is the case for the second problem
when k = 2 and Ω is bounded (see [60] and [89]).
In Theorem C, we actually prove that the optimal partition, for the minimization
of the addition of the energies of a system of two harmonic oscillators in R2 , is com-
posed of the supports of the positive and negative parts of its second eigenfunction.
The free boundary problem, which usually arises for competitive system when g12
goes to infinity, has been studied by Caffarelli and Lin in [51], for the singularly
perturbed elliptic system −∆u + g12 u2 v = 0 and −∆v + g12 v 2 u = 0. They prove
the existence of a limiting pair, for which each component is a harmonic function
in its domain of definition. They also prove the C 0,α regularity of the nodal line
{u = v = 0}. They use Algrem’s monotonicity formula, originally established in
[20], which allows to prove Hölder uniform estimates for the solutions. See [49], [59]
and [140] for other works in this spirit.
A further related problem is the study of (3.1.3) in the attractive case when g1 and
g2 are both negative. This is certainly a very different problem, since in this case
the associated energy may not be bounded from below (see for example [121]) and
no ground states exist. Symmetry results have been proved in this case, when there
is neither trapping potential nor mass constraints. In [164], Wei and Weth prove
that in a system defined over all R2 , they are infinitely many non radial solutions,
and in [166], that in a system defined in the unit ball, for every k ∈ N∗ there is a
radially symmetric solution (u, v) such that u − v changes exactly k times of sign
in the radial variable. We also mention the work of Liu in [123], where he does not
prescribe the constraint of the L2 -norm of each component, and therefore, in the
limit only one component remains. This is another form of segregation, since the
model allows one component to disappear in the strong interacting case.
Sketch of proofs
We now explain the outline of the paper and the main ideas of the proofs.
In Section 3.2, we prove Theorem A and the segregation of the minimizing pairs
(ug , vg ). We first remark that for all g ∈ R3+ and (u, v) ∈ Y , Eg (u, v) = Eg1 ,g2 ,∞ (u, v),
so Eg (ug , vg ) ≤ Eg1 ,g2 ,∞ (u, v), and there are positive constants c1 , c2 and c3 such that
for every g1 ≥ 0 and g2 ≥ 0
Eg (ug , vg ) ≤ c1 + c2 g1 + c3 g2 , (3.1.9)
for all g12 ≥ 0.
Sketch of the proof of Theorem A : Using (3.1.9), the energy of (ugn , vgn ) (and also
the Lagrange multipliers λgn and µgn ) is uniformly bounded. Hence, there exists
(u∞ , v∞ ) ∈ X which is the weak limit of (ugn , vgn ) in HV1 × HV1 and the strong limit
in L2 × L2 . We prove that (u∞ , v∞ ) minimizes Eg1 ,g2 ,∞ in Y using energy estimates.
0
In order to prove the Cloc convergence, we follow the ideas Wei and Weth in [165],
where they prove the equicontinuity of solutions of system (3.1.3) in bounded do-
mains and without trapping potential. We show that the proof of Theorem 1.1(a)
therein works in our setting. The reason of this, is that the proof consists in a resca-
ling of the solutions, which yields in a limit problem over all R2 . Rescaling ugn and
vgn identically, we get the same limit problem, so the equicontinuity holds, which
gives the local uniform convergence. To show that (u∞ , v∞ ) satisfies the system
(3.1.7), we first prove some estimates for minimizing pairs (ug , vg ) of Eg : there are
√
C2 > 0 and C3 > 0 such that, kug k∞ , kvg k∞ ≤ C2 and k∇ug k∞ , k∇vg k∞ ≤ C3 g12 ,
for any g ∈ [0, c0 ]2 × R+ with g12 large enough. Using these, we show in Proposition
−η
3.3 that there is C4 > 0, such that for every ε > 0 and every η > 1, |vg | ≤ C4 g12 in
−η
{inf g ug > ε} and |ug | ≤ C4 g12 in {inf g vg > ε}, for g12 large enough depending on ε
n
and η. This implies that the functions g12 |vgn |2 ugn and g12
n
|ugn |2 vgn converge weakly
to zero respectively in {|u∞ | > 0} and {|v∞ | > 0}, and the system (3.1.7) is satisfied.
Remark 1.2. We expect the derivatives of u∞ and v∞ to be not continuous trough

the nodal line {u∞ = v∞ = 0}, and the system (3.1.7) to be solved only in the
interior of the supports of u∞ and v∞ . In [110] and [126], we find simulations sup-
porting this idea. In addition, in the case when g1 = g2 , we conjecture that the
difference of the two components is a smooth function. Indeed, as we explain in Re-
mark 3.6, in this case, we expect u∞ − v∞ to solve a homogeneous elliptic equation,
which after standard elliptic regularity arguments, yields in smoothness for u∞ −v∞ .
In Section 3.3, we study the properties of fully segregated BECs. We show that the
minimizers of Eg1 ,g2 ,∞ in Y are locally Lipschitz continuous, and that the nodal set
has empty interior. The local Lipschitz continuity is an important result because la-
ter in the proof of Theorem C, we need the limiting function to be locally Lipschitz
continuous. The result about the nodal line of the segregated minimizers is used
in the proof of Theorem C. We stress here that the space Y is not a manifold, so
we cannot perform calculus of variations therein. We have to use other techniques
to deal with the minimizers of Eg1 ,g2 ,∞ in Y , in order to get a system of equations
allowing us to study their local properties. We will explain these techniques, based
on the works of Conti, Terracini and Verzini in [58] and [60].
In Section 3.4, we prove Theorem C and Theorem B. Theorem B follows directly

from Theorem A and Theorem C(ii). The key ingredient in the proof of Theorem C
is the extremal property of the half-spaces Ha,ν = {x · ν > a} with respect to the
Gaussian-Rayleigh quotient
R
|∇u|2 dµ
F (u) = R .
|u|2 dµ
Here a is a real number, ν ∈ S 1 and µ is the Gaussian measure over R2 . In [71],
Ehrhard prove, roughly speaking, that if Λ(S) is the infimum of F over the functions
vanishing outside of a Lebesgue measurable set S, and Ha,ν is a half-space with the
same Gaussian measure as S, then
Λ(S) ≥ Λ(Ha,ν ) ,
and the equality holds only when S = Ha,ν for some ν ∈ S 1 .
Sketch of the proof of Theorem

√ C : For the first assertion, we start noticing that the
1 2
change of variables ũ(x) = 2π α u(αx)e 4 |x| , whit α = (2ω)−1/2 , gives
E0,0,∞ (u, v) = ω (F (ũ) + F (ṽ) + 2) .

Next, we first show in Lemma 3.7 that if (u0 , v0 ) minimize E0,0,∞ over Y , then
F (ũ0 ) = Λ(Ũ0 ) and F (ũ0 ) = Λ(Ṽ0 )

where Ũ0 and Ṽ0 are respectively the supports of ũ0 and ṽ0 . Using Ehrhard’s result,
we establish then that
Ũ0 = Ha,ν and Ṽ0 = Hb,ν 0

for some a, b ∈ R2 and ν, ν 0 ∈ S 1 . Because Ũ0 and Ṽ0 are disjoint, and because
the nodal line has empty interior (see Proposition 3.6), we get that a = −b and
ν 0 = ν. Hence Ṽ0 = Ha,ν c
. Then, we use a result of Beckner, Kenig and Pipher
(see [48]) saying that the mapping a 7→ Λ(Ha,ν ) is convex, to prove that a is equal
to zero. The definition of the change of variables gives then that supp u0 = H0,ν
c
and supp v0 = H0,ν for some ν ∈ S 1 . Finally, using standard arguments for one-
component BECs, we prove that system (3.1.7) is uniquely solved in half-spaces.
The second assertion of Theorem C comes then easily, remarking that the positive
and the negative parts of a second eigenfunction of the harmonic oscillator solve
(3.1.7) in half-spaces.
3.2 The segregation limit

In this section we assume that g1 and g2 are bounded with respect to g12 , so there
is c0 > 0 such that
max{ g1 , g2 } ≤ c0 . (3.2.1)
Moreover, with out loss of generality, we assume that ug and vg are real positive
functions over all R2 . We are allowed to do this after the following result, which is
standard for one component BECs (see [4]) :
Lemma 3.1. Each component of a minimizing pair of Eg over X is, up to a complex

multiplier of modulus one, a real positive smooth function.
Proof. Every minimizing pair (ug , vg ) solves the system (3.1.3), so using standard
elliptic regularity and the strong maximum principle, ug and vg are non vanishing
smooth complex functions. Thus, there are smooth real functions ϕ1 and ϕ2 , such
that ug = |ug |eiϕ1 and vg = |vg |eiϕ2 . The diamagnetic inequality imply that (|ug |, |vg |)
is also a minimizer of Eg over X. We have then the equality Eg (|ug |, |vg |) = Eg (ug , vg ),
which imply that ϕ1 and ϕ2 are constants, and hence the result.
We start showing uniform estimates on Eg and on ug , vg and its derivatives. We will

use these to prove Theorem A and the segregation of ug and vg in Proposition 3.3.
Estimates on minimizers
Lemma 3.2. There are positive constants C0 , C1 , C2 and C3 , such that if (ug , vg )
is a minimizing pair of Eg over X, with g = (g1 , g2 , g12 ) ∈ [0, c0 ] × [0, c0 ] × R+ , and
λg , µg are the associated Lagrange multipliers, then
Eg (ug , vg ) ≤ C0 (3.2.2)
0 < λg , µg ≤ C1 (3.2.3)
kug k∞ , kvg k∞ ≤ C2 (3.2.4)
√
k∇ug k∞ , k∇vg k∞ ≤ C3 g12 (3.2.5)
for every g12 > 1.
Proof. Proof of (3.2.2) : as we saw in (3.1.9), if (u, v) ∈ Y then
Eg (ug , vg ) ≤ Eg (u, v) = Eg1 ,g2 ,∞ (u, v) ,

so there are positive constant c0 , c1 and c2 , not depending on g1 , g2 or g12 , such that
for every g1 ≥ 0 and g2 ≥ 0
Eg (ug , vg ) ≤ c0 + c1 g1 + c2 g2 = C0 ,
for all g12 ≥ 0. We get then (3.2.2) after (3.2.1).
3.2. THE SEGREGATION LIMIT 75
Proof of (3.2.3) : Multiplying the first equation in (3.1.3) by ug and then integrating
over all R2 we get that
Z
λg = |∇ug |2 + V u2g + g1 u4g + g12 u2g vg2 ,
so after the mass constraint λg > 0, and after (3.2.2), λg ≤ 4 Eg (ug , vg ) ≤ 4 C0 = C1 .

The same argument is valid with µg , which yields (3.2.3).
Proof of (3.2.4) : Consider x ∈ R2 and R > 0. Using (3.2.2), the mass constraint
and the continuous embedding H 1 ,→ Lp for p ∈ [2, ∞), for every ball B = B2R (x)
0 0
there is a positive constant C = C (p, R) such that
q
0 0 0
p
kug kLp (B) ≤ C kug kH 1 (B) ≤ C 2 Eg (ug , vg ) + 1 ≤ C 2 C0 + 1 . (3.2.6)
After (3.1.3) we get
−∆ug ≤ hg in B
00 00
with hg = λg ug . Using (3.2.3) together with (3.2.6), there is C = C (q, R, g1 , g2 ) > 0
such that
00
khg kLq/2 (B) ≤ C (3.2.7)
for every q ∈ [4, ∞).
Using a local estimate for H 1 subsolutions of elliptic equations (see Theorem 8.17
000 000
in [80]) there is C = C (R, p, q) > 0 such that
000
R− /p ku+ 2−4/q
2

sup ug ≤ C g kLp (B) + R khg kLq/2 (B) .
BR (x)
Fixing R, p and q, we derive from (3.2.6) and (3.2.7) that there is C2 > 0 such that
ug (x) ≤ C2
for every g12 > 0. The same argument is valid with vg . Therefore, since ug and vg
are positive, we get (3.2.4).
Proof of (3.2.5) : we first prove that ug and vg have polynomial decay at infinity.
More precisely, we claim that for every α > 0 there is rα > 0 and Cα > 0 such that
for all g12 ≥ 0,
Cα
ug (x) < (3.2.8)
|x|α
p
for all x ∈ Kα = R2 \ Brα (0). For α fixed, take rα2 = 1/2ω2 C1 + C12 + 4α2 ω 2 . A
straightforward calculation shows that
α
rα
fα (x) = C2
|x|
is a supersolution of the first equation in (3.1.3),
−∆fα + fα (V + g1 fα2 + g12 |v|2 − λg ) ≥ 0 in Kα

while

fα = C2 ≥ ug .

∂Kα ∂Kα
Now define ψ = fα − ug and suppose that ψ is strictly negative somewhere in Kα .

Because fα and ug are of class C 2 in Kα and go both to zero at infinity, ψ must have
a local minimum x0 in K̊α : ug (x0 ) > fα (x0 ) and D2 ψ(x0 ) is positive defined, so
∆ψ(x0 ) ≥ 0. Using this, and the fact that fα is a super solution of the first equation
in (3.1.3) while ug is a solution, we have
(V (x0 )+g1 fα (x0 )2 +g12 vg (x0 )2 −λg )fα (x0 ) ≥ (V (x0 )+g1 ug (x0 )2 +g12 vg (x0 )2 −λg )ug (x0 ) .
But our choice of rα implies in particular that V (x0 ) − λg > 0, so we get a contra-
diction with ψ(x0 ) < 0 and the claim is proved. Remark that the previous claim,
together with (3.2.4) imply that V ug is uniformly bounded in R2 with respect to g12 .
To finish the proof, let x ∈ R2 and suppose that g12 > 1. For y ∈ B2 (0) define
−1/2 −1/2
ũ(y) = u(x + g12 y) and ṽ(x) = v(x0 + g12 y). We have
−1 −1/2
∆ũ(y) = g12 {V (x + g12 y)ũ(y) + g1 ũ(y)3 − λg ũ(y)} + ũ(y)ṽ 2 (y) ,
so after (3.2.1), (3.2.3), (3.2.4) and the claim there is a constant c > 0 such that
|∆ũ(y)| ≤ c for all y ∈ B2 (0) and g12 > 1. Using a Hölder estimate for the first
derivative of ũ (see Theorem 8.32 in [80]) there is a constant C > 0 such that
k∇ũkL∞ (B1 (0)) ≤ C(C2 + c) .
We get then the result with C3 = C(C2 + c) considering ∇ũ (0).
We now show the key ingredient in the Proof of Theorem A, the segregation of ug
and vg . This is a generalization of Proposition 2.1 in [53], to positive solutions of
(3.1.3) defined in all R2 . We have also used in the proof some ideas from [165].
Proposition 3.3. Let (ug , vg ) be a sequence of minimizers of Eg over X and define

n o n o
Uε ≡ x ∈ R2 ; inf ug (x) ≥ ε , Vε ≡ x ∈ R2 ; inf vg (x) ≥ ε .
g12 >0 g12 >0
For any ε > 0 and η > 1, there are G0 > 0 and a positive constant C4 such that
−η −η
vg ≤ C4 g12 in Uε and ug ≤ C4 g12 in Vε
for every g12 > G0 .
Proof. Let η > 1, ε > 0 and x ∈ Uε . Define the numbers

2
η1 = 8 η ρ ∈ (0, 21 e−C2 )
−1/2 −ρ
sg = η1 g12 ln g12 tg = g12
so sg < tg for g12 large enough depending on η and ε, and note B = Bsg (x).
Define also the function hg : (0, ∞) → R by

Z
1
hg (r) = u2 ds ,
2πr ∂Br (x) g
and notice that
Z
2πr h0g (r) = ug · ∂ν ug ds . (3.2.9)
∂Br (x)
After (3.2.4) we have that 0 < hg ≤ C22 . We claim the existence of ξg ∈ (sg , tg ) such
that
−1
h0g (ξg ) ≤ . (3.2.10)
ξg ln ξg
If not, we will get
tg
−1
Z
ln sg
C22 > hg (tg ) − hg (sg ) > dr = ln
sg r ln r ln tg
ln (η1 ln g12 ) − 12 ln g12 g12 →∞

1
= ln −−−−→ ln ( ) ,
−ρ ln g12 2ρ
which contradicts the choice of ρ.
We have then
Z Z Z Z
2 2
|∇ug | dx ≤ |∇ug | dx = ug · ∂ν ug ds − ∆ug · ug dx .
B Bξg (x) ∂Bξg (x) Bξg (x)
The first term of the right hand side can be estimated using (3.2.10), and the second
one using (3.1.3), (3.2.3) and (3.2.4). We get
Z
2π
|∇ug |2 dx ≤ − + C1 C22 ξg2 ,
B ln ξg
0
so there is a positive C such that
Z 0
C
|∇ug |2 dx ≤ (3.2.11)
B ln g12
for g12 large enough.
Using Theorem 7.17 in [80], we have that for p ∈ (2, 3) and γ = 1 − p2 , there is
Cp > 0 such that
2/p
oscB ug ≤ Cp sγg k∇ug k(p−2)/p
∞ k∇ug k2 .
00
Using (3.2.5), (3.2.11) and the definition of sg , there is C > 0 such that
00 − γ + 12 − p2 3
oscB ug ≤ C η1γ g122 (ln g12 )1− p
00 3
= C η1γ (ln g12 )1− p ,
so after (3.2.1), oscB ug → 0 when g12 → ∞. This implies that

ug ≥ in B
2
for g12 large enough.
Using this last estimate, together with (3.1.3) (3.2.3) and (3.2.4), we get
2


 −∆vg ≤ −g12 16 vg in B




vg ≥ 0 in B

(3.2.12)


vg ≤ C2 in ∂B





for g12 large enough. Hence, Lemma 4.4 in [58] gives that exist a constant C > 0
(not depending in g, ε, η or x), such that
q
sg 2
− g12 16
kvg k∞ ≤ C C2 e 2 in B sg (x) .
2
Tacking C4 = C C2 , the definition of sg gives

−η
vg (x) ≤ C4 g12 ,
for g12 large enough depending on η and ε. The equivalent argument holds for ug in
Vε , which yields Proposition 3.4.
We have now all the tools to prove Theorem A.
Proof of Theorem A. Let (un , vn ) = (ugn , vgn ) be a sequence of minimizing pairs

of Egn in X with g1n → g1 , g2n → g2 and g12
n
→ ∞.
(i) After (3.2.2) the sequences un and vn are bounded in HV1 , so there exists
(u∞ , v∞ ) ∈ HV1 × HV1 with (up to a subsequence)
un * u∞ in HV1
vn * v∞ in HV1 ,
as n → ∞. The compact embedding HV1 (R2 ) ,→ L2 (R2 ) (see Lemma 2.1 in [100])
gives the strong L2 convergence, so ku∞ k2 = kv∞ k2 = 1 and (u∞ , v∞ ) ∈ X.
On the one hand, we have that
kun vn − u∞ v∞ k1 ≤ kun k2 kvn − v∞ k2 + kv∞ k2 kun − u∞ k2 = o(1) , (3.2.13)

so (up to a subsequence) un vn → u∞ v∞ a.e. in R2 . And on the other hand, after
(3.2.4) and (3.2.8), un vn ≤ C1 h ∈ L2 . The Lebesgue dominated convergence theorem
together with (3.2.2) gives then
C2
ku∞ v∞ k2 = lim kun vn k2 ≤ lim n
= 0.
n→∞ n→∞ g12
Hence, u∞ · v∞ = 0 a.e. in R2 and (u∞ , v∞ ) ∈ Y .
0
In order to prove the Cloc convergence, we follow directly the ideas of Wei and Weth
in [165]. In Theorem 1.1(a) therein, they show that sequences of positive solutions of
a class of competitive nonlinear elliptic systems in bounded domains are uniformly
equicontinous. Their proof consists in a rescaling of the solutions and the domains,
which yields in a limit problem over all R2 . We will show that defining the rescaled
functions of ugn and vgn , we get the same limit problem, so the equicontinuity holds.
We recall that (un , vn ) satisfies the system (3.1.3) and that after Lemma 2.2, the
sequence is uniformly bounded in HV1 × HV1 and in L∞ × L∞ .
Following the proof of Theorem 1.1(a) in [165], if the sequence (un , vn ) is not uni-
formly equicontinuous, there exists δ > 0 such that, without loss of generality, un
satisfies (up to a subsequence)
inf |x − y| ; x, y ∈ R2 , |un (x) − un (y)| ≥ 2δ → 0

as n → ∞.
Then, since the un functions are positives, (3.2.8) implies that there are xn , yn ∈ R2
such that rn := |xn − yn | → 0 as n → ∞, dn := un (yn ) ≥ δ and un (xn ) ≥ dn + δ.
Take e1 = (1, 0) and choose An ∈ O(2) such that An e1 = rn−1 (yn − xn ). We define
the rescaled function vi,n : R2 → R+ by
v1,n (x) = un (xn + rn An y) and v2,n (x) = vn (xn + rn An y) .

Then, v1,n solves in R2


 −∆v1,n = l1,n v1,n − rn2 g12
n
v1,n
v1,n > 0

(3.2.14)

 v 1,n 1 )
(e = dn ≥ δ
v1,n (0) ≥ v1,n (e1 ) + δ .


Here l1,n (x) = rn2 V (xn + rn An x) + g1n v1,n
2
(x) − λn , which after (3.2.8), (3.2.3) and
(3.2.4) satisfies
l1,n v1,n → 0 in L∞ as n → ∞ . (3.2.15)

Moreover, after (3.2.2) v1,n is uniformly bounded in H 1 , and Lemma 3.3 also applied
for the sequence (v1,n , v2,n ). This last two properties, together (3.2.14), (3.2.15), im-
plies that (v1,n , v2,n ) satisfies the same hypotheses as in the proof of Theorem 1.1(a)
in [165]. Hence, we obtain the same limit problem when n goes to infinity, which
following exactly the proof, yields a contradiction. The desired result then holds.
(ii) For the first assertion, let (ũ, ṽ) be any pair in Y . Then
Eg1n ,g2n ,∞ (ũ, ṽ) = Egn (ũ, ṽ) ≥ Egn (un , vn ) ≥ Eg1n ,g2n ,∞ (un , vn ) . (3.2.16)
Since the pair (un , vn ) satisfies the uniform bounds (3.2.8) and (3.2.4), the L2 conver-
gence implies the L4 convergence. This, together with (3.2.16) and the weak lower
semicontinuity of the HV1 norm, gives
Eg1 ,g2 ,∞ (ũ, ṽ) = lim inf Eg1n ,g2n ,∞ (ũ, ṽ)
n→∞
≥ lim inf Eg1n ,g2n ,∞ (un , vn )
n→∞
≥ Eg1 ,g2 ,∞ (u∞ , v∞ ) ,
which imply the result since after (i), (u∞ , v∞ ) ∈ Y .
For the second assertion, let ϕ be a C ∞ function supported in K ⊂⊂ {u∞ > 0}.
Multiplying the first equation on (3.1.3) by ϕ and then integrating, we get
Z Z
n 3 n 2
∇ϕ · ∇un + ϕ (V un + g1 un − λn un ) = − ϕ g12 vn un . (3.2.17)
K K
Using the weak convergence of un to u∞ , the left hand side of (3.2.17) tends to
Z
∇ϕ · ∇u∞ + ϕ (V u∞ + g1 u3∞ − λu∞ )
with λ the limit (up to a subsequence) of λn , which exists because of (3.2.3).
After (i), un converges uniformly to u∞ in K. Hence, K ⊂⊂ {u∞ > 2ε} for some
ε > 0, and there exists N > 0 such that
K ⊂ { inf ugn ≥ ε} .
n>N
n 2
Thus, Proposition 3.3, together with (3.2.4), implies that g12 vn un converges uni-
formly to zero in K, so the right hand side of (3.2.17) tends to zero as n → ∞.
Hence,
Z Z
3
∇ϕ · ∇u∞ + ϕ (V u∞ + g1 u∞ ) = λ ϕu∞ .
The same argument is valid with v∞ , which yields the result.

3.3. PROPERTIES OF FULLY SEGREGATED COMPONENTS 81
3.3 Properties of fully segregated two-component

BECs
In this section we prove some properties of fully segregated two-component BECs.
The results in here will be used to prove the symmetry breaking in the limit case
when g1 = g2 = 0. We start with the local Lipschitz continuity of minimizers of
Eg1 ,g2 ,∞ in Y .
Local Lipschitz continuity

Proposition 3.4. If (u, v) is a nonnegative real minimizer Eg1 ,g2 ,∞ over Y , then u
and v are locally Lipschitz continuous in R2 .
To prove this proposition, we first see in Lemma 3.5 that each component of a mini-
mizing pair of Eg1 ,g2 ,∞ is a weak subsolution of a Gross-Pitaevskii equation, and that
the difference of both components is a weak solution of a non homogeneous elliptic
equation. We derive local uniform estimates with respect to g1 and g2 for the mi-
nimizing pairs. These estimates imply that the minimizers satisfy some extremality
conditions. We conclude using the following result of Conti, Terracini and Verzini
([60], Theorem 8) :
Theorem 3.1. (Conti-Terracini-Verzini) Let Ω be a bounded regular set of R2 ,
M ≥ 0 and w1 , w2 ∈ H 1 (Ω) such that w1 ≥ 0, w2 ≥ 0 and w1 · w2 = 0. If w1 and w2
satisfy
−∆w1 ≤ M , −∆w2 ≤ M and − M ≤ −∆(w1 − w2 ) ≤ M ,

then they are both Lipschitz continuous in the interior of Ω.
Lemma 3.5. Let g1 ≥ 0, g2 ≥ 0 and (u, v) a nonnegative minimizer of Eg1 ,g2 ,∞ over
Y . Then,
− ∆u + V u + g1 u3 ≤ λu , − ∆v + V v + g2 v 3 ≤ µv (3.3.1)
and
− ∆(u − v) + V (u − v) + g1 u3 − g2 v 3 = λu − µv (3.3.2)
weakly in R2 . Here λ = e1 (u) and µ = e2 (v), where
Z
ei (w) = |∇w|2 + V |w|2 + gi |w|4 , i = 1, 2 .
Proof. Proof of (3.3.1) : Arguing by contradiction, suppose that

Z
∇u · ∇φ + (V u + g1 u3 − λu) φ > 0
for some 0 ≤ φ ∈ C0∞ (R2 ).
For t ∈ (0, 1), define a new test function as :

(u − tφ)+

(w1 , w2 ) = ,v .
k(u − tφ)+ k2
Where u+ = max(u, 0) and u− = max(−u, 0). In the rest of the proof the o(·) nota-
tion will mean with respect to the t → 0 limit.
Since {(u − tφ)+ > 0} ⊂ {u > 0}, (u − tφ)+ · v = 0 a.e. in R2 , and (w1 , w2 ) ∈ Y .
Using f 2 = (f + )2 + (f − )2 , kuk2 = 1 and 0 ≤ (u − tφ)− ≤ tφ, we compute
Z
tφ)+ k22 [(u − tφ)+ ]2 − u2

k(u − = 1+
Z
2tφu + [(u − tφ)− ]2 − t2 φ2

= 1−
Z
= 1− 2tφu + o(t) ,
so
Z Z
1 1
=1+ 2tφu + o(t) and =1+ 4tφu + o(t) .
k(u − tφ)+ k22 k(u − tφ)+ k42
The difference between the energies is then
Z Z Z
1 + 2 2
tφu · |∇(u − tφ)+ |2

Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) = |∇(u − tφ) | − |∇u| +
2
Z Z Z
1 + 2 2
tφu · V (x)|(u − tφ)+(3.3.3)
|2

+ V [(u − tφ) ] − u +
2
Z Z Z
1 + 4 4
tφu · g1 |(u − tφ)+ |4 + o(t) .

+ g1 [(u − tφ) ] − u +
4
We note that
Z Z
1 1
|∇(u − tφ)+ |2 − |∇u|2 ≤ |∇(u − tφ)|2 − |∇u|2

2 2
Z
= −t ∇φ · ∇u + o(t) ,
so (3.3.3) becomes
Z
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) ≤ −t ∇u · ∇φ + V uφ + g1 u3 φ − e1 ((u − tφ)+ ) uφ + o(t) .
Using Lebesgue dominated convergence theorem, we have that e1 ((u−tφ)+ ) → e1 (u)

when t → 0, so
3.3. PROPERTIES OF FULLY SEGREGATED COMPONENTS 83
Z
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) ≤ −t ∇u · ∇φ + (V u + g1 u3 − λu) φ + o(t) .
Hence, for t small enough we get the contradiction
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) < 0 .

Using the same arguments with v and µ, the inequalities in (3.3.1) are proved.
Proof of (3.3.2) : Define û = u − v and suppose that

Z
∇û · ∇φ + (V û + g1 u3 − g2 v 3 − λu + νv)φ < 0
for some 0 ≤ φ ∈ C0∞ (R2 ).
For t ∈ (0, 1), define a new test function as :
(û + tφ)+ (û + tφ)−

(w1 , w2 ) = , .
k(û + tφ)+ k2 k(û + tφ)− k2
As before, we compute
Z Z
1 1
=1− 2tφu + o(t) , =1− 4tφu + o(t) ,
k(û + tφ)+ k22 k(u + tφ)+ k42
Z Z
1 1
=1+ 2tφv + o(t) and =1+ 4tφv + o(t) .
k(û + tφ)− k22 k(u + tφ)− k42
Using that u · v = 0, the difference between the energies is
Z
1
|∇(û + tφ)|2 − |∇û|2 + V |(û + tφ)|2 − |û|2

Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) =
2
Z
1
g1 |(û + tφ)+ |4 − |u|4 + g2 |(û + tφ)− |4 − |v|4

+
4
Z
u e1 ((û + tφ)+ ) − v e2 ((û + tφ)− ) φ + o(t) .

−t
Hence,
Z
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) = t ∇û · ∇φ + V ûφ + (g1 u3 − g2 v 3 )φ
Z
u e1 ((û + tφ)+ ) − v e2 ((û + tφ)− ) φ + o(t) .

− t
Using the same argument as before, we see that e1 ((û + tφ)+ ) − e1 (u) = o(1) and
e2 ((û + tφ)− ) − e2 (v) = o(1), so
Z
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) = t ∇û · ∇φ + (V û + g1 u3 − g2 v 3 ) , φ
Z

− t u λ − v µ φ + o(t) .
And again, for t small enough we get
Eg1 ,g2 ,∞ (w1 , w2 ) − Eg1 ,g2 ,∞ (u, v) < 0 ,

a contradiction.
We have proved the inequality

Z
∇(u − v) · ∇φ + V (u − v)φ + (g1 u3 − g2 v 3 )φ − (λu − ν∞ v)φ ≥ 0 .
Using the same arguments with v̂ = v − u we get

Z
∇(v − u) · ∇φ + V (v − u)φ + (g2 v 3 − g1 u3 )φ − (−λu + ν∞ v)φ ≥ 0
for every 0 ≤ φ ∈ C0∞ (R2 ). Equality (3.3.2) is then proved.
Proof of Proposition 3.4 : Let Ω be any bounded set of R2 . After (3.3.1), u and
v are respectively H 1 subsolutions of −∆u = λu and −∆v = µv. Arguing as in the
proof of (3.2.4) u and v are uniformly bounded in Ω with respect to g1 and g2 , and
since they minimize Eg1 ,g2 ,∞ , λ and µ are also uniformly bounded. There are then
positive M1 , M2 = O(g1 , g2 ) such that
−∆u ≤ M1 and − ∆v ≤ M2 in Ω ,
so after (3.3.2) and the previous estimates, there are positive M3 , M4 = O(g1 , g2 )
such that
−M3 ≤ −∆(u − v) ≤ M4 in Ω .
Theorem 3.1 with M = max{M1 , M2 , M3 , M4 } implies then that u and v are Lip-
schitz continuous in Ω and the result is proved.
The nodal set

We now prove that the nodal set has empty interior. We do this by following an idea
of Chang et all. in [53]. The point is that if the nodal set has a ball B contained in
it, then u can be extended to a solution ũ of an elliptic equation in supp u ∪ B, but
then, after the strong maximum principle and the mass constraint, ũ cannot vanish
in the interior of its support.
3.4. THE NON INTERACTING LIMIT 85
Proposition 3.6. If (u, v) is a nonnegative real minimizer of Eg1 ,g2 ,∞ over Y , then
the nodal set {x ∈ R2 ; u(x) = v(x) = 0} has no interior points.
Proof. After Proposition 3.4, we know that u and v are locally Lipschitz functions,
so U = {x ∈ R2 ; u(x) > 0} and V = {x ∈ R2 ; v(x) > 0} are open regular sets.
Define Ũ = R2 \ V and suppose that the nodal line has an interior point. Then
U ( Ũ and the Lebesgue measure of U is less than the Lebesgue measure of Ũ. Let
w̃ be the minimizer of
Z
1 1 1
Eg1 ,0,0 (w) = |∇w|2 + V |w|2 + g1 |w|4
Ũ 2 2 4
R
over the functions w in H01 (Ũ) such that Ũ w2 = 1.
Is clear that (w̃, v) ∈ Y , so after Eg1 ,g2 ,∞ (w̃, v) ≥ Eg1 ,g2 ,∞ (u, v) we get Eg1 ,0,0 (u) ≤
Eg1 ,0,0 (w̃). This imply that u solves −∆u + (V + g1 |u|2 − λ)u = 0, with λ defined as
in Lemma 3.5. Therefore, by elliptic regularity, u is a C 2 function in Ũ, and using
the strong maximum principle, u ≡ 0 because it vanish in the interior of Ũ. This
contradicts the mass constraint, so the nodal line has no interior points.
Remark 3.6. As we said in the introduction, we expect the derivatives of u and v

to be not continuous trough the nodal line. In the case when g1 = g2 , we also expect
the difference of the two components to be a smooth function. Indeed, after Pro-
position 3.5, if (u, v) is a nonnegative real minimizer of Eg1 ,g2 ,∞ over Y , then u − v
solves the elliptic equation (3.3.2) in R2 . When g1 = g2 , we expect the Lagrange
multipliers λ and µ to be equal, and equation (3.3.2) to be homogeneous. Standard
elliptic regularity theory implies then that u − v is a C ∞ function over all R2 . These
properties are verified when g1 and g2 are both equal to zero : in Theorem C(ii) we
show that in this case, u and v are respectively the positive and negative parts of a
second eigenfunction of the harmonic oscillator L in R2 , so u − v ∈ C ∞ (R2 ), λ = µ,
and there is a jump of the derivatives of u and v trough the nodal line.
3.4 The non interacting limit

In this section we study the minimizers (u0 , v0 ) of E0,0,∞ over Y . We prove Theorem
C, that is, the two components are supported in complementary half-spaces meeting
at zero, and that they are respectively the positive and negative parts of a second
eigenfunction of the harmonic oscillator −∆ + ω 2 |x|2 in R2 .
For a ∈ R and ν ∈ S 1 we define the half space
Ha,ν = {x ∈ R2 ; x · ν > a} .
We write Ha = Ha,(0,1) . The main idea in the proof of Theorem C is the extremal
property of half-spaces with respect to the Gaussian-Rayleigh quotient
R
|∇f |2 dµ
F (f ) = R .
|f |2 dµ
Here µ is the Gaussian measure in R2 , which density is
1 − 1 |x|2
dµ(x) = e 2 dx .
2π
We remark that the invariance of the Gaussian measure with respect to rotations
gives that for every ν, ν 0 ∈ S 1
c
Λ(Ha,ν 0 ) = Λ(Ha,ν ) = Λ(H−a,ν ). (3.4.1)
Moreover, since R̄ 3 a 7→ µ(Ha,ν ) is an increasing function, for every Lebesgue mea-
surable set S there is a real a such that µ(S) = µ(Ha,ν ).
For a non empty open S, we define F(S) as the class of nonnegative non zero
functions absolutely continuous on lines with support in S, and Λ by
Λ(S) = inf F (f ) .
f ∈F (S)
In [71] and [72], Ehrhard studies isoperimetric inequalities in Gauss spaces and
introduce the Gaussian symmetrization, a variant of classical symmetrizations used
to solve isoperimetric problems, such as the principal frequency of a membrane or
the torsional rigidity of a bar (see for example [46], [151] or [158]). In [72], Ehrhard
proves that among all subsets with prescribed Gaussian measure, half-spaces have
minimal Λ :
Theorem 3.2. (Ehrhard) Let S be a non empty open subset of R2 and a ∈ R such
that µ(S) = µ(Ha ). Then,
Λ(S) ≥ Λ(Ha )
and the equality holds if and only if S = Ha,ν for some ν ∈ S 1 . Moreover, the
infimum in Λ(Ha ) is attained by some f ∈ F(Ha ).
We recall that the trapping potential is given by (3.1.2), so after (3.1.4) the energy
writes
Z n o 1Z n
1 2 2 2 2 2 2 2 2
o
E0,0,∞ (u, v) = |∇u| + ω |x| |u| + |∇v| + ω |x| |v| ,
2 R2 2 R2
and that the class of minimization Y is given by (3.1.5).
To prove Theorem C, we first perform a change of variable in order to deal with the
minimization problem in a different setting. For (u, v) ∈ Y define
√ 1 2 √ 1 2
ũ(x) = 2π α u(αx) e 4 |x| and ṽ(x) = 2π α v(αx) e 4 |x| (3.4.2)
with α = (2ω)−1/2 , and
Ỹ = {(ũ, ṽ) ; (u, v) ∈ Y } .

A straightforward computation gives

E0,0,∞ (u, v) = ω F (ũ) + F (ṽ) + 2 . (3.4.3)
Lemma 3.7. A nonnegative pair (u, v) minimizes E0,0,∞ over Y if and only if (ũ, ṽ)
minimize F (ũ) + F (ṽ) over Ỹ . Moreover, in this case ũ minimizes F over F(Ũ) and
ṽ minimizes F over F(Ṽ), where Ũ = supp ũ and Ṽ = supp ṽ.
Proof. The first assertion is immediate after (3.4.3). If (u, v) is a nonnegative mini-
0,1
mizer of E0,0,∞ over Y , then after Proposition 3.4 u is in Cloc (R2 ), so it is absolutely
continuous on lines, and because of the mass constraint u is not identically zero.
After (3.4.2) the same properties hold for ũ, so ũ ∈ F(Ũ). For every w̃ ∈ F(Ũ) with
finite Gaussian-Rayleigh quotient, (w/kwk2 , v) ∈ Y . Thus, (3.4.3) gives F (ũ) ≤ F (w̃).
The same argument is valid with ṽ, so the second assertion is proved.
We are now able to prove Theorem C. For the first part, we use the ideas of Beckner,
Kenig and Pipher in Section 2.4 of [48].
Proof of Theorem C(i). To lighten the notation we write u = u0 , U = supp u,

Ũ = supp ũ and the analogous for v0 . The diamagnetic inequality implies that
(|u|, |v|) is also a minimizer of E0,0,∞ over Y , so we suppose, without loss of ge-
nerality, that u and v are nonnegative real functions.
Step 1 : Ũ = Ha,ν and Ṽ = Hb,ν 0 . Suppose that Ũ or Ṽ is not a half-space. Then, after
Lemma 3.7 and Theorem 3.2 there are real numbers a, b such that µ(Ũ) = µ(Ha ),
µ(Ṽ) = µ(Hb ) and
F (ũ) + F (ṽ) = Λ(Ũ) + Λ(Ṽ) > Λ(Ha ) + Λ(Hb ) . (3.4.4)

We claim that a + b ≥ 0. First, since µ(Ũ ∩ Ṽ) = 0 we have
µ(Ha ) + µ(Hb ) ≤ 1 , (3.4.5)

which implies that a and b cannot be both negative. We suppose then without loss
of generality that
b<0≤a and a + b < 0. (3.4.6)

c
The fact that µ(Ha ) + µ(H−a ) = 1, together with (3.4.5), implies that µ(Hb ) ≤
µ(H−a ), which contradicts (3.4.6). The claim is then proved.
c
The inequality a + b ≥ 0 implies that Ha ∩ H−b = ∅. Hence, for every pair (w̃1 , w̃2 ) ∈
c
F(Ha ) × F(H−b ), ( /kw1 k2 , /kw2 k2 ) ∈ Ỹ . After (3.4.3) and (3.4.4) we obtain
w̃1 w̃2
F (w̃1 ) + F (w̃2 ) ≥ F (ũ) + F (ṽ) > Λ(Ha ) + Λ(Hb ) .

Minimizing F in the previous inequality with respect to w̃1 ∈ F(Ha ) and with res-
c
pect to w̃2 ∈ F(H−b ), and considering (3.4.1), we obtain the contradiction Λ(Ha ) +
Λ(Hb ) > Λ(Ha ) + Λ(Hb ), so Step 1 is proved.
Step 2 : a = −b and ν = ν 0 . Since ũ · ṽ = 0, µ(Ha,ν ∩ H−b,ν 0 ) = 0, which imply

that the boundaries of Ha,ν and H−b,ν 0 must be parallel, i.e., ν = ν 0 and a ≥ −b.
Moreover, if a > b, then the nodal set {ũ = ṽ = 0} has a non empty interior, which
after (3.4.2) contradicts Proposition 3.6. We have shown that
c
Ũ = Ha,ν and Ṽ = Ha,ν
for some a ∈ R, ν ∈ S 1 .
Step 3 : a = 0. First, the monotonicity of the first eigenvalue of the Dirichlet problem
with respect to the domain gives that a 7→ Λ(Ha ) is an increasing function. Moreover,
the same argument of Theorem 2.4.5 in [48] (see also Theorem 6.2 in [43]) gives that
it is a convex function. Considering (3.4.1) we derive
Λ(Ha ) + Λ(Hac ) Λ(Ha ) + Λ(H−a ) Λ(H0 ) + Λ(H0c )

= ≥ Λ(H0 ) = . (3.4.7)
2 2 2
Suppose now that a 6= 0 and consider (w̃1 , w̃2 ) ∈ F(H0 )×F(H0c ). The same argument
used in Step 1, together with (3.4.7), gives
F (w̃1 ) + F (w̃2 ) > Λ(Ha ) + Λ(Hac ) ≥ Λ(H0 ) + Λ(H0c ) .

After Theorem 3.2, the infima in Λ(H0 ) and Λ(H0c ) are attained, so again, minimizing
F in the previous inequality with respect to w̃1 ∈ F(H0 ) and with respect to w̃2 ∈
F(H0c ), we obtain the contradiction Λ(H0 ) + Λ(H0c ) > Λ(H0 ) + Λ(H0c ). We have
proved that
c
Ũ = H0,ν and Ṽ = H0,ν ,
which considering (3.4.2), gives
c
U = H0,ν and V = H0,ν
for some ν ∈ S 1 .
We now give the proof of the second part √ ofiθTheorem C, that is, every minimi-
+ iθ− −
zing pair of E0,0,∞ over Y is of the form 2 (e wν , e wν ) for some ν ∈ S 1 and
+
θ+ , θ− ∈ R.
We recall that the second eigenvalue of the harmonic oscillator −∆ + ω 2 |x|2 over
R2 has multiplicity 2. An orthogonal base, with respect to the L2 product, of the
associated spectral space is given by
ω|x|2 ω|x|2
η1 (x) = cω x1 e− 2 and η2 (x) = cω x2 e− 2
p
with cω = 2/π ω. Thus, every function in the spectral space, with L2 norm equal
to 1, is the of the form
ω|x|2
wν (x) = cω (x · ν) e− 2 (3.4.8)
for some ν ∈ S 1 .
Proof of Theorem C(ii). We write u0 = u, v0 = v and H0,ν = Hν . After Theorem

A(ii) and Theorem C(i), there is ν ∈ S 1 such that u solves

 −∆u + ω 2 |x|2 u = λ u in Hν
R u = 0 in ∂Hν (3.4.9)
2
u = 1

Hν
R
with λ = Hν |∇u|2 + ω 2 |x|2 u2 , and v solves the equivalent problem in Hνc .
The same argument as in Lemma 3.1 imply that u = eiθ+ |u| in Hν and v = eiθ− |v| in
Hν , with |u| and |v| positive and θ+ , θ− ∈ R. We claim now that there is uniqueness
for the modulus of u in problem (3.4.9). Suppose that there are two positive solutions
u1 and u2 of (3.4.9) respectively with λ1 and λ2 . Suppose that λ1 ≤ λ2 and define
h = u1 /u2 in Hν . We have that
∇(u22 ∇h) = (λ2 − λ1 ) u22 h , (3.4.10)

and using the mass constraint that
Z Z
2 1
u2 h(h − 1) = u22 (h − 1)2 .
Hν 2 Hν
Multiplying (3.4.10) by h−1, then integrating over Hν and performing an integration

by parts we find
Z
1
u22 |∇h|2 + (λ2 − λ1 ) u22 (h − 1)2 = 0 ,
Hν 2
which imply that h ≡ cte. The mass constraints imply then that h ≡ 1, so the claim
is proved. The equivalent result is valid for v in Hνc .
Finally, a direct computation shows that

√ ω|x|2 √ ω|x|2
uν (x) = 2 cω (x · ν)+ e− 2 and vν (x) = 2 cω (x · ν)− e− 2
are positive solutions of problem (3.4.9) respectively in Hν and Hνc , so u = eiθ+ uν
and v = eiθ− vν for some ν ∈ S 1 and θ+ , θ− ∈ R.
We finish this article by given the proof of Theorem B.
Proof of Theorem B. Let (ugn , vgn ) be a sequence of minimizing pairs of Egn , with
g1n → 0, g2n → 0 and g12
n
→ ∞. After Theorem A, ugn (respectively vgn ) converges
locally uniformly to u0 (respectively v0 ), where (u0 , v0 ) minimizes E0,0,∞ over Y .
Theorem C(ii) gives then that u0 = eiθ+ wν+ and v0 = eiθ− wν− for some ν ∈ S 1 and
θ+ , θ− ∈ R.
The author would like to thank C. Kenig, J. Wei, A. Aftalion and S. Terracini.
Chapitre 4
A minimal interface problem arising

from a two component Bose-Einstein
condensate via Γ-convergence
Amandine Aftalion 1 and Jimena Royo-Letelier 2 3
Article soumis
Abstract
We consider the energy modeling a two component Bose-Einstein conden-
sate. In the limit of strong coupling and strong segregation, we prove the
Γ-convergence to a perimeter minimization problem, with a weight given by
the density of the condensate. In the case of equal mass for the two compo-
nents, this leads to symmetry breaking for the ground state. The proof relies
on a new formulation of the problem in terms of the total density and spin
functions, which turns the energy into the sum of two weighted Cahn-Hilliard
energies. Then, we use techniques coming from geometric measure theory to
construct upper and lower bounds. In particular, we make use of the slicing
technique introduced in [22].
4.1 Introduction
The aim of this paper is to prove a Γ-convergence result for a functional modeling
a two component Bose-Einstein condensate in the case of segregation. We introduce
a new formulation of the problem which transforms the two wave functions descri-
bing each component of the condensate into total density and spin functions. The
new functional in the density and spin variables is given by the sum of two weighted
1. CNRS et Université Versailles-Saint-Quentin-en-Yvelines, Laboratoire de Mathématiques de
Versailles, CNRS UMR 8100, 45 avenue des États-Unis, 78035 Versailles Cédex, France
2. Laboratoire de Mathématiques de Versailles, CNRS UMR 8100, 45 avenue des États-Unis,
78035 Versailles Cédex, France.
3. Ceremade, CNRS UMR 7534, Université Paris-Dauphine, Place du Maréchal de Lattre de
Tassigny, 75775 Paris Cédex 16, France.
91
92 CHAPITRE 4. A PHASE TRANSITION PROBLEM
Cahn-Hilliard energies modeling phase transition problems as in the Modica-Mortola

problem [130]. In fact, our new functional is strongly related to that of Ambrosio-
Tortorelli approaching the Mumford-Shah image segmentation functional [22]. We
use techniques coming from geometric measure theory [17, 18, 22, 40] to construct
upper and lower bounds for our initial functional and prove Γ-convergence to a per-
imeter minimization problem, with a weight given by the density of the condensate.
There is a large mathematical literature about the segregation patterns for two com-
ponent Bose Einstein condensates [31, 33, 51, 59, 140, 165] : regularity of the limiting
functions, regularity of the interface, asymptotic behaviour near the interface. All
these papers use the limiting equations and do not take into account the trapping
potentials and the Γ convergence of the energy as we do.
Before introducing the functional for a two component Bose Einstein condensate,
we recall some properties of a single Bose Einstein condensate (BEC). A single BEC
is described by the wave function η minimizing the energy
Z
1 1 1
Eε (η) = |∇η|2 + 2 V (x)|η|2 + 2 |η|4 (4.1.1)
2 R2 ε 2ε
where V is the trapping potential, usually taken to be harmonic, that is V (x) =

|x|2 , ε is a small parameter giving rise to a large coupling constant describing the
self interaction
R of the condensate. The minimization is performed under the mass
constraint R2 |η|2 = 1. The ground state is (up to multiplication by a constant) a
real positive function. Let
Z
2 2
ρ(x) = max(λ − |x| , 0) with λ > 0 chosen such that ρ = 1 where D = B(0, λ).
D
(4.1.2)
√
Then, when ε is small, the ground state ηε is close to the function ρ in D, with
exponential decay at infinity. Properties of ηε can be found in [4, 12, 77, 100, 108].
A two component Bose Einstein condensate can be experimentally realized as 2
isotopes of the same atom in different spin states [84] or isotopes of different atoms
[133]. They are described by two wave functions u1 and u2 , respectively representing
components 1 and 2. The Gross Pitaevskii energy of the two component condensate
is given by Z
1
Eε (u1 , u2 ) = Eε (u1 ) + Eε (u2 ) + gε |u1 |2 |u2 |2 , (4.1.3)
2 R 2
where Eε is given by (4.1.1) and gε is the intercomponent coupling strength. The

energy is minimized under the mass constraints
Z
|uj |2 = αj with αj > 0 and α1 + α2 = 1 . (4.1.4)
R2
In [126], numerical simulations have been performed to classify the ground states
according to the values of ε, gε and also the rotational velocity. For ε small and gε
large, the numerical evidence is that, for α1 = α2 = 1/2, the preferred ground state
is such that each component is asymptotically located in a half disk with a local
inverted parabola profile. If α1 6= α2 , they occupy sections in a disk, the area of which
is proportional to αi . In particular, when neither αi is too small, this configuration
has less energy than a disk vs annulus configuration, which also provides segregation
but preserves symmetry. Observation of symmetry breaking has also been obtained
experimentally very recently [128]. The breaking of symmetry has been analyzed in
[153] in a different limit, namely in the case ε large and gε large.
Here, we assume strong coupling between components, that is, gε → ∞, and we
study the regime
gε ε2 → +∞ and ε → 0. (4.1.5)
A trick introduced in [126] is to use a spin formulation also called the nonlinear
sigma model. In our special setting, since the ground states are non vanishing real
functions (up to multiplication by a complex number of modulus 1), this amounts
to defining
p !
|u1 |2 + |u2 |2 ϕ |u1 | + i|u2 |
v := and := Arg p , (4.1.6)
ηε 2 |u1 |2 + |u2 |2
where ηε is the ground state of (4.1.1) with L2 norm 1. The definition of ϕ implies
that |u1 |2 − |u2 |2 = ηε2 v 2 cos ϕ. The mass constraints (4.1.4) rewrite as
Z Z
2 2
ηε v = α1 + α2 = 1 and ηε2 v 2 cos ϕ = α1 − α2 . (4.1.7)
R2 R2
We point out that cos ϕ corresponds to the third component of the spin function.
Because there is no rotation in the system, the ground states are, up to multipli-
cation by a complex number of modulus one, positive functions. Thus, the second
component of the spin is zero and the first one is sin ϕ.
Since the components are expected to segregate, the expected behaviour is thus that
v tends to 1 except on a transition line corresponding to the interface between the
two components, while ϕ tends to 0 on component 1 and π on component 2. This
is what we want to analyze rigorously.
We split the energy into its main contributions and will prove that
Eε (u1 , u2 ) = Eε (ηε ) + Fε (v) + Gε (v, ϕ) (4.1.8)
where Eε is given by (4.1.1), ηε is the ground state of Eε and

Z
1 1
Fε (v) = ηε2 |∇v|2 + 2 ηε4 {1 − v 2 }2 , (4.1.9)
2 R2 2ε
Z
1
Gε (v, ϕ) = ηε2 v 2 |∇ϕ|2 + ηε4 v 4 g̃ε {1 − cos2 (ϕ)} (4.1.10)
8 R2

and g̃ε = gε 1 − gε1ε2 . Since ηε2 converges to ρ given by (4.1.2) in D, the limits of
Fε and Gε can be analyzed as the limits of
Z
1 1
ρ|∇vε |2 + 2 ρ2 {1 − vε2 }2 (4.1.11)
2 D 2ε
Z
1
ρv 2 |∇ϕε |2 + ρ2 vε4 g̃ε {1 − cos2 (ϕε )}. (4.1.12)
8 D ε
These two energies are of Modica-Mortola types with a weight which vanishes on
the boundary of D. Given the definition of ϕε , there is a domain where cos ϕε tends
to 1 (asymptotic region of component 1) and a domain where cos ϕε tends to −1

(asymptotic region of component 2), and thus a transition region exists between the
two domains. Two options exist for vε :
– either vε goes to√1 everywhere, which makes the first energy small and the second
energy of order g̃ε ,
– or vε goes to zero on the transition line where cos ϕ varies from +1 to −1 : this
makes the second energy of lower order and the first energy of order C/ε.
Because of our hypothesis that ε2 g̃ε tends to infinity, it is the second scenario which
costs less energy. Though vε goes to 1 on each component, it has a transition region
of size ε where it goes sharply to zero. The second energy is of lower order and
cannot be seen in the limit. It has just the effect of creating a small region around
the interface where vε is small. The first energy can be analyzed with techniques
coming from [22] and, once the rescaling in ε is made, the Γ-limit comes from the
problem on lines :
Z ∞
1 0 2 1 2 2 2
I(x) = inf ρ(x)(w ) + ρ(x) (1 − w ) ; w ∈ Lip(R+ ) , w(0) = 0 and w(+∞) = 1 .
2 0 2
Using the Euler-Lagrange equation associated with I, we see that for x ∈ D, the
infimum is attained by the function
r !
ρ(x)
wx (t) = tanh t ,
2
and we have
Z 1
3/2 1
I(x) = σρ(x) with σ = √ {1 − t2 } dt . (4.1.13)
2 0
This means that wx is the optimal profile transition at the point x, and that σρ(x)3/2
is the minimum energy needed by w, to go from 0 to 1 at x. In the limiting energy,
we have 2σρ(x)3/2 because as ε → 0, vε goes from 1 to 0 on one side of the interface
between the two components, and from 0 to 1 on the other side. Therefore, we expect
the limit to be defined as the integral on the interface where ϕ goes from 0 to π
of the function 2σρ(x)3/2 . This requires a precise mathematical definition for this
interface.
We define X as the space of functions ϕ ∈ BVloc (D ; {0, π}) such that
Z
ρ cos ϕ = α1 − α2 . (4.1.14)
R2
We will prove the Γ-convergence of ε(Eε (·, ·) − Eε (ηε )) to F given in X by

Z
2σ
ρ /2 |Dϕ| .
3
F(ϕ) =
π D
The limiting energy F measures the length, with a weight of ρ3/2 , of the interface
between the two phases of ϕ. Each phase of ϕ corresponds to one component of
the totally segregated two-component limiting condensate. Notice that when F(ϕ)
is finite, {ϕ = π} has finite perimeter in compact subsets of D, and
Z Z
3/2 1
ρ /2 dH1 .
3
F(ϕ) = 2σ ρ dH = 2σ
D ∩ ∂ ∗ {ϕ=π} D ∩ Sϕ
Here ∂ ∗ {ϕ = π} stands for the reduced boundary of {ϕ = π} and Sϕ is the comple-

ment of the Lebesgue points of ϕ, that is,
Z
1
Sϕ = x ∈ D ; @ t ∈ R such that lim+ 2 |ϕ(y) − t| dy = 0 .
r→0 πr Br (x)
We refer to [21, 74, 82] for the geometric measure theory concepts. We also refer to
[17] for an introduction to the theory of Γ-convergence and to the Modica-Mortola
theorem by G. Alberti.
We now state our main theorem :
Theorem 4.1. Let us assume that V (x) = |x|2 , and let

Z
1 2 1 2 2 2
H = (u1 , u2 ) ∈ H (R ; R) × H (R ; R) , V (u1 + u2 ) < ∞ , (u1 , u2 ) satisfies (4.1.4) .
R2
The functional ε(Eε (·, ·) − Eε (ηε )) Γ-converges with respect to the L1loc (D) × L1loc (D)
distance to F(ϕ), in the following sense :
(Compactness) for every sequence {(u1,ε , u2,ε )}ε>0 of minimizers of Eε in H such
that
sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) < +∞ , (4.1.15)

ε>0
there exists ϕ ∈ X and a (not relabeled) subsequence such that
√
L1loc (D) × L1loc (D) ; (4.1.16)

(u1,ε , u2,ε ) → ρ 1{ϕ=0} , 1{ϕ=π} in
and (Lower bound inequality)

ε→0
(Upper bound inequality) For every ϕ ∈ X, there exists a sequence {(u1,ε , u2,ε )}ε>0 ⊂
√
H, converging as ε → 0 to ρ 1{ϕ=0} , 1{ϕ=π} in L1loc (D) × L1loc (D), such that
lim sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) ≤ F(ϕ) . (4.1.18)

ε→0
We point out that we only prove the Γ-convergence at the level of minimizers of
Eε . Indeed, minimizers of the functional have the property that they are positive
functions which do not vanish. Therefore, this property allows the definition of (v, ϕ)
through (4.1.6). As usual, the Γ-convergence theorem implies the convergence of the
energy of the ground states :
Corollary 4.2. If {(u1,ε , u2,ε )}ε>0 is a sequence of minimizer of Eε in H, then
lim ε (Eε (u1,ε , u2,ε ) − Eε (ηε ) ) = inf F . (4.1.19)

ε→0 X
Corollary 4.3. If α1 = α2 = 1/2, then for ε sufficiently small, the solutions

(u1,ε , u2,ε ) are not radial.
Remark 4.4. Our main theorem remains true when V is any trapping potential
for which we have good estimates for the ground state ηε , namely the estimates in
Proposition 4.2.
4.1.1 Links with related problems

The segregation behaviour in two component condensates has been widely studied :
regularity of the wave function [59, 140, 165], regularity of the interface [51], asymp-
totic behaviour near the interface [31, 33]. The main difference with these references
is that, on the one hand, we use mainly the energy instead of the equation and,
on the other hand, we do not switch off the trapping potential by blowing up the
problem near the interface or by considering a bounded domain with no trapping.
Moreover, in the strong interaction regime, it is the trapping potential which pro-
vides the leading order behaviour of the wave function through the inverted parabola
profile ρ. In all the previous quoted references, the trapping potential is not present,
whereas we just deal with it by a proper division of the limiting wave function which
allows to express nicely the energy using a trick introduced by [114].
Nevertheless, we believe that our approach could bring information on an estimate
left open in [31] : that gε m4ε is bounded, where mε is the value where u1 and u2 get
equal, which in our case is probably related to the minimum of vε . We detail this
remark in section 5.3.
4.1.2 Main ideas in the proof

Let us now give more details on the proof.
The proof consists of upper and lower bounds, that we construct for the functional
Fε (vε , ϕε ) = ε (Eε (u1,ε , u2,ε ) − Eε (ηε ) ).
√
For the upper bound, we choose the set A where asymptotically u2 will be ρ. In
a first step, we assume that ϕ = π1A , where A is an open bounded subset of R2
with smooth boundary such that H1 (∂A ∩ ∂D) = 0. The test function ϕε is matched
between 0 in a subdomain of D \ Ā to π in a subdomain of A, using a transition
region of size εtε . In order to approximate the optimal 1 dimensional profile that
solves I(y), we define


 mε in (0, tε )
tanh in (tε , T )

wε,T =

 h in (T, T + 1/T )
1 in (T + 1/T , +∞),

where tε = tanh mε and h is a polynomial which matches smoothly tanh to 1. Then

we define r !
y ρ(y)
wε,T (t) = wε,T t ,
2
for t = d(x)/ε < CT , and d(x) is the distance to the boundary. In order to construct
yi
vε , we need a partition of unity for ∂A, where we match the functions wε,T , as yi
varies along this partition. For this vε , we can estimate Fε with techniques similar to
those of Modica-Mortola [130], and to the adaptation of these techniques to problems
with weight by Bouchitté [40]. Because ρ vanishes, we cannot use directly the results
of Bouchitté and we need precise estimates on the behaviour of ηε near the boundary.
Since wε,T is the optimal profile for the 1D of (4.1.11), there is a transition
R from 1 to
0 and a transition from 0 to 1, we find an upper bound which is 2 ∂A I(y) dy. Then
we prove that for this test function, Gε (vε , ϕε ) is lower order : indeed the transition
layer for ϕ is is of order εtε , so much smaller than the one of vε . Hence in Gε , vε can
be approximated by mε . We choose m4ε = ε2 gε , which tends to 0, and makes Gε of
lower order.
This provides the upper bound for an open bounded subset A with smooth boundary
such that H1 (∂A ∩ ∂D) = 0. We show in the appendix that for any ϕ ∈ X, {ϕ = π}
can be approximated by sets A which are open bounded subsets of R2 with smooth
boundary such that H1 (∂A ∩ ∂D) = 0 and that the mass constraints can be satisfied
for the approximating u1,ε , u2,ε .
The difficulty in the lower bound is to prove that vε goes to zero on a line and
that it provides a positive lower bound. Indeed, the usual Modica-Mortola bound
would imply that vε goes to 1 almost everywhere and the lower bound is 0. We
have to use Gε and the upper bound to prove that vε has a transition to 0 and that
cos2 ϕε tends to 1. Hence, because of the mass constraint, we get two regions where
asymptotically ϕε is 0 and π. To analyze the behaviour of vε , we use the slicing
method introduced in [22] or [42]. This consists in looking at the transition for vε
in one dimension, perpendicular to the interface and get the 1D energy estimate.
The use of the energy Gε is only to prove that vε goes to zero. We first prove the
lower bound for εFε in 1D using the coarea formula, and then in 2D using the slicing
method. We get that εFε (vε , ϕε , E) converges to a measure µ(E) supported in Sϕ of
density ρ3/2 with respect to the H1 measure. The last part of the proof of the lower
bound is inspired by the ideas in [18].
We end with a variant of the coarea formula that can be found in [125] Lemma 2.2,
and in [40] Proposition 2.
Proposition 4.1. Let Ω be an open bounded subset of RN , and Ψ(x, s, p) a Borel
function of Ω × R × RN , which is sublinear in p. Let u be a Lipschitz continuous
function on Ω and denote, for every t > 0, St = {x ∈ Ω ; u(x) < t}. Then, for
almost every t ∈ R , 1St belongs to BV (Ω) and we have
Z Z ∞ Z
Ψ(x, u, Du) dx = dt Ψ(x, t, D1St ) . (4.1.20)
Ω −∞ Ω
The paper is organized as follows : in section 2, we present the properties of ηε .
Then in section 3, we prove the decoupling of energy (4.1.8) and how to go from the
(u1 , u2 ) formulation to (v, ϕ). Section 4 is devoted to the upper bound, and section
5 to the lower bound. Finally, in section 6, we prove our main theorem.
4.1.3 To go further
Analysis of the limiting problem
A natural question is to analyze the limiting problem, that is the ground state of
F under the constraint (4.1.14). If we define A to be the set where cos ϕ = 1. Then
R R
ρ = α1 and D\A ρ = α2 . We do not have any rigorous result for the limiting
A
problem. In particular, we have not proved that the only limiting domains are either
disks sectors or a disk and annulus.
If ρ = 1, then the problem of minimizing F amounts to minimizing |∂A| under
the constraints |A| = α1 and |D \ A| = α2 . The Euler-Lagrange equation of the
minimization problem yields that the curvature is either 0 or constant, hence A is
either a disk, an annulus or a disk sector. The equivalent problem with a weight ρ
is open.
If we assume that the solution is either two disks sectors or a disk and an annulus,
we can compute explicitly the energy F and find that if α1 = α2 , then the optimal
configuration is two half disks, while if α1 is much less then α2 , then the ground state
is a disk and an annulus (see section 6.4). Indeed, the energy of two √ disk sectors is
3/4 1/2
3σ/2, while the energy of a disk and annulus is 8σ(1 − α1 ) (1 − 1 − α1 ) if α1
corresponds to the mass of the inside disk. If α1 or α2 = 1 − α1 is to small, then the
disk and annulus becomes the preferred configuration. In the case α1 = α2 = 1/2,
it follows from our theorem that symmetry breaking occurs since at the limit, the
disk plus annulus configuration does not minimize the energy. These two cases are
well illustrated in the experimental observations of [128], figure 4.
We insist on the point that a rigorous analysis of the ground states of F in X is an
interesting open question.
Convergence for u1,ε , u2,ε

√
The convergence that we have for u1,ε , u2,ε to ρ(1{ϕ=0} , 1{ϕ=π} ) is very weak.
Nevertheless, we expect that on compact subsets of 1{ϕ=π} or 1{ϕ=0} , the convergence
can be improved. For instance, it would be natural to have similar convergence as
√ 1
that of ηε to ρ (that is Cloc ) on these domains.
Case gε ε2 of order 1
An interesting open question is to deal with the case when gε ε2 tends to a positive
finite constant c20 . In this case, Fε and Gε become of the same order and we expect
that m = lim inf vε is a positive constant (on the interface where ϕ varies), instead
of being 0. We believe that our techniques still provide an upper bound for the
problem. We expect the Γ-limit to be
Z
π 3 1
ρ /2 |Dϕ| .
3
2σm + c0 m
4 π D
R1
where σm = √1 (1 − t2 ) dt.
2 m
Case of different scattering lengths

Another related case is when the two components are different atoms with different
scattering lengths which leads to energies Eε which depends on the component
namely Z
1 1 gi
Eε,i (η) = |∇η|2 + 2 |x|2 |η|2 + 2 |η|4
2 R2 ε 2ε
4.2. ESTIMATES FOR ηε 99
with g1 6= g2 . Then the leading order Thomas Fermi approximation is no longer the
same for each component, namely it is
gi ρi = λ2i − |x|2 in Bi = B(0, λi ).
The limiting problem becomes R : find a partition of B1 ∪ B2 into three sets A1 , A2

and N , such that u2i,ε → 1Ai , Ai ρi = αi and it minimizes
Z Z
2 g1 2 g2
|x| ρ1 + ρ1 + |x|2 ρ2 + ρ22 . (4.1.21)
A1 2 A2 2
This problem is open and is probably related to the problem of finding a partition
of the disk into two subdomains which minimize the sum of the first eigenvalues of
the Dirichlet laplacian.
Of course, in our case, since we have B1 = B2 , ρ1 = ρ2 and N = ∅, (4.1.21) does
not provide any information at leading order. This is why we have to go to the next
order which yields the perimeter minimization problem.
4.2 Estimates for ηε

Let ηε be the ground state of Eε given by (4.1.1), with L2 norm 1. The ground state
is a non vanishing radially symmetric function. It is unique up to multiplication by
a constant of modulus one, and satisfies the Gross-Pitaevskii equation
1 2 1 λε
− ∆ηε + 2
|x| ηε + 2 |ηε |2 ηε = 2 ηε . (4.2.1)
ε ε ε
The term ε−2 λε is the Lagrange multiplier associated with the mass constraint,
and the pair (ηε , λε ) is unique among positive solutions of (4.2.1). As ε tends to
√
0, ηε tends to ρ given by (4.1.2). Throughout the paper, we will need precise
estimates for this convergence. The following proposition, based on previous results
in [12, 76, 77, 100, 108], sums up the properties of ηε . We point out that it follows
√
from [76, 77, 108] that an approximation of ηε by ρ holds as close to the boundary
as needed and is given by (4.2.5). We also include an estimate of ρ in terms of the
distance to the bulk that will be used in the proofs.
Proposition 4.2. There are constants c, C > 0, α ∈ (1/2, 3/5) and γ ∈ (1/2, 3/4), such
that for ε sufficiently small, ρ, λ being given by (4.1.2),
Eε (ηε ) ≤ C/ε2 , (4.2.2)

C ε | ln ε| /2 ,
1
|λε − λ| ≤ (4.2.3)
√
kηε − ρkC 1 (K) ≤ CK ε2 | ln ε| for K ⊂⊂ D , (4.2.4)
√
|ηε (x) − ρ(x)| ≤ C εγ for x ∈ B(0 , λ − c εα ) , (4.2.5)
−1/3 (λ−|x|)
ηε (x) ≤ C ε /6 ec ε for x ∈ R2 \D ,
1
(4.2.6)
∂r ηε (|x|) ≤ 0 for x ∈ R2 , (4.2.7)
ρ(x)
∈ [1, 2) for x ∈ D . (4.2.8)
λ dist(x, ∂D)
Proof : for the proof of (4.2.2), one can rewrite the energy as
Z
1 1 2 1 2
Eε (η) = Eε (η) + 2 λ − ρ (4.2.9)
2ε 2 D
where
Z
1 1 2 1
Eε1 (η) = |∇η|2 + 2
|η|2
− ρ(x) + 2 (λ2 − |x|2 )− |η|2 ,
2 R2 2ε ε
and (λ2 − |x|2 )− is the negative part of (λ2 − |x|2 ). In Theorem 2.1 of [12], it is
proved
R 2 that Eε1 (η) ≤ C| ln ε|. Then (4.2.2) follows from (4.2.9) and the fact that
D
ρ = 2λ2 /3.
Estimate (4.2.4) is proved in Proposition 2.2 of [100]. Estimates (4.2.3) and (4.2.6)
are proved in Theorem 2.1 of [12]. Estimate (4.2.7) is also proved in Theorem 2.1 of
[12], but only in a neighborhood of ∂D. But the proof, however, works in the case
V (x) = |x|2 and the estimate holds in all R2 .
We now prove (4.2.5). For λ > 0, we define η̃ε,λ as the unique radially symmetric,
positive solution of the equation
− ε2 ∆η + (λ2 − |x|2 )η + η 3 = 0 . (4.2.10)

The function η̃ε,λ corresponds to a ground state of a BEC without mass constraint.
In [76, 77, 100], the behavior of η̃ε,λ is studied. Using the results in Proposition 1.2,
Remark 1.3 and Proposition 1.4 in [76], we obtain
1 − |x|2

1/3
η̃ε,1 (x) = ε ν0 + O(ε) ,
ε2/3
where
1 5
ν0 (y) = y /2 − y − /2 + Oy→+∞ (y − /2 ) , y ∈ (−∞, ε− /3 ] .
1 11 2
2
Hence, for x ∈ B(0, 1) we obtain
p
1 − |x|2 | ≤ C ε2 (1 − |x|2 )− /2 + ε4 (1 − |x|2 )− /2 + ε .
5 11

| η̃ε,1 (x) −
In particular, if x ∈ B(0 , λ − εα ) with α ∈ (1/2, 3/5), we get
p
1 − |x|2 | ≤ C ε2 ε− + ε4 ε−
5α/2 11α/2
+ ε = O(εγ )

| η̃ε,1 (x) − (4.2.11)
with γ ∈ (1/2, 3/4). We will use (4.2.11) to prove (4.2.5). First, a straight computation
shows that defining ελ = λ−2 ε, η̃ελ ,λ solves equation (4.2.10) with λ = 1. Hence,
considering (4.2.11), a change of variables gives
√
| η̃ελ ,λ (x) − ρ(x) | = O(εγ ) , (4.2.12)
for x ∈ B(0 , λ − (λ−2 ε)α ). In Proposition 2.2 and Theorem 2.2 in [100], it is proved
that
k∇ηε kL∞ (R2 ) = O(ε−1 ) ; (4.2.13)

4.3. REWRITING THE ENERGY 101
and that
1/2
ηε,λ (x) = `ε,λ η̃ε̃,λ (`−1
ε,λ x) , (4.2.14)
where

ελε
`ε,λ = 1+ and ε̃ = `−1
ε,λ ε .
λ
It follows from (4.2.3) that
`ε,λ = 1 + O(ε2 | ln ε| /2 ) ε̃ = ε + O(ε2 | ln ε| /2 ) .

1 1
and
Hence, using (4.2.13) and (4.2.14), we obtain
ηε,λ (x) = η̃ε̃,λ (x) + O(ε | ln ε| /2 ) .

1
Putting this last estimate in (4.2.12), and using that γ ∈ (1/2, 3/4), we obtain that
√
| ηελ (x),λ − ρ(x) | = O(εγ ) ,
for x ∈ B(0 , λ − c εα ) with c > 0. We obtain (4.2.5) by changing ελ by ε in the
previous estimate. Finally, writing
ρ(x) (λ + |x|)
=
λ dist(x, ∂D) λ
we get (4.2.8) for |x| < λ.
4.3 Rewriting the energy

In this section, we prove equality (4.1.8), that is, the reformulation of the Gross-
Pitaevskii energy of a two component condensate in (4.1.3), as the weighted Cahn-
Hilliard energy for the pair (v, ϕ) defined by (4.1.6), plus the energy of the ground
state ηε of a one component condensate. We start by giving the properties of the
minimizers of Eε and the properties of the corresponding pairs (vε , ϕε ) defined by
(4.1.6).
Proposition 4.3. (i) Let {(u1,ε , u2,ε )}ε>0 be a sequence of minimizing pairs of Eε
in H satisfying (4.1.15). Then, each component is a non vanishing smooth function,
and there is C > 0 such that
ku1,ε kL∞ (R2 ) , ku2,ε kL∞ (R2 ) < C (4.3.1)

for every ε > 0. Moreover, the pairs (vε , ϕε ) are well defined by (4.1.6), verify the
mass constraints (4.1.7) and we have
(vε , ϕε ) ∈ Liploc (R2 ; (0, +∞) × [0, π]) (4.3.2)

and
sup kvε kL∞ (K) < CK for every K ⊂⊂ D . (4.3.3)
ε>0
(ii) Conversely, let (v, ϕ) ∈ Lip(R2 ; (0, +∞) × [0, π]) satisfying (4.1.7) such that
v , ∇v , ∇ϕ ∈ L∞ (R2 ). Then, defining
u1 = ηε v cos (ϕ/2) and u2 = ηε v sin (ϕ/2) , (4.3.4)

we have (u1 , u2 ) ∈ H and |u1 |2 + |u2 |2 > 0.
Proof : (i) Let (u1,ε , u2,ε ) be a minimizer of Eε in H. Since Eε (|u1,ε |, |u2,ε |) ≤
Eε (u1,ε , u2,ε ), the pair of the absolute values satisfies the system
− ∆u1,ε + ε−2 V + ε−2 u21,ε + gε u22,ε u1,ε = λ1,ε u1,ε

(4.3.5)
−∆u2,ε + ε−2 V + ε−2 u22,ε + gε u21,ε u2,ε = λ2,ε u2,ε ,

(4.3.6)
where λ1,ε and λ2,ε are the Lagrange multipliers associated with (4.1.4). The strong
maximum principle yields that |u1,ε | and |u2,ε | are positive functions. Using standard
elliptic regularity, we deduce further that u1,ε and u2,ε are non vanishing smooth
functions. We use an argument in [100] to prove that u1,ε and u2,ε are uniformly
1/2
bounded in R2 . Let us define w = ε−1 |u1,ε | − λε . We have w ∈ L3loc (R2 ) and
∆w ∈ L1loc (R2 ). Kato’s inequality and equation (4.3.5) give
∆(w+ ) ≥ sgn+ (w) ∆w ≥ ε−3 sgn+ (w) εw (εw + ελε/2 ) (εw + 2ελε/2 ) ≥ (w+ )3 .
1 1
Hence, −∆(w+ ) + (w+ )3 ≤ 0 weakly in R2 and Lemma 2 in [44] yield w+ ≤ 0.

1/2
We obtain |u1,ε | ≤ ελε . Multiplying equation (4.3.5) by u1,ε and then integrating
1/2 p
we find λε ≤ 2 Eε (|u1,ε |, |u2,ε |). Since (u1,ε , u2,ε ) verifies (4.1.15), from estimate
(4.2.2), we derive
q
0 < |u1,ε | ≤ ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) + Eε (ηε ) < C .
We similarly prove that 0 < |u2,ε | < C, so (4.3.1) is proved. Since ηε > 0, u1,ε and
u2,ε do not vanish in R2 , the pairs (vε , ϕε ) are well defined by (4.1.6) and vε > 0.
Since u1,ε and u2,ε are smooth, vε and ϕε are locally Lipschitz functions so (4.3.2)
holds. The definition of vε and (4.1.4) give
Z
ηε2 vε2 = α1 + α2 .
R2
From the definition of ϕε , we infer that
|u1,ε |2 − |u2,ε |2
cos(ϕε ) = ,
|u1,ε |2 + |u2,ε |2
which, together with (4.1.4), yields
Z
ηε2 vε2 cos ϕε = α1 − α2 .
R2
Hence, (vε , ϕε ) satisfies (4.1.7). Finally, the estimate (4.2.4) gives ηε ≥ cK > 0 in
K ⊂⊂ D, so (4.3.1) yields (4.3.3).
4.3. REWRITING THE ENERGY 103
(ii) Consider (v, ϕ) as in the statement and define (u1 , u2 ) by (4.3.4). Since (v, ϕ)
verifies (4.1.7), relation (4.3.4) gives
Z Z
2 2
|u1 | + |u2 | = ηε2 v 2 = α1 + α2
R2 R2
and
Z Z
2 2
|u1 | − |u2 | = ηε2 v 2 cos2 ϕ = α1 − α2 .
R2 R2
Thus, (u1 , u2 ) verifies (4.1.4). We have |u1 |2 +|u2 |2 > 0. Indeed, if it was not the case,
since v > 0 then ϕ should take simultaneously the values 0 and π. Since v ∈ L∞ (R2 ),
bounds (4.2.5) and (4.2.6) on ηε give
Z Z
2 2
V ηε2 < +∞ .

V |u1 | + |u2 | ≤ C
R2 R2
We compute
|∇u1 |2 ≤ C v 2 |∇ηε |2 + ηε2 |∇ηε |2 + v 2 ηε2 |∇ϕ|2 .

The right hand side of the inequality is integrable in R2 because v , ∇v , ∇ϕ ∈

L∞ (R2 ) and ηε ∈ H 1 (R2 ) ∩ L∞ (R2 ). Thus, u1,ε ∈ H 1 (R2 ). We prove similarly that
u2,ε ∈ H 1 (R2 ). We have proved that (u1,ε , u2,ε ) ∈ H.
We now prove the rewriting of the energy.
Proposition 4.4. Let (u1 , u2 ) ∈ H satisfying |u1 |2 + |u2 |2 > 0. Defining (v, ϕ) by
(4.1.6) we have
Eε (u1 , u2 ) = Eε (ηε ) + Fε (v) + Gε (v, ϕ) ,
where Eε , Fε and Gε are given respectively by (4.1.1), (4.1.9) and (4.1.10).
Proof : since |u1 |2 + |u2 |2 > 0, the pair (v, ϕ) is well defined. The definitions of v
and ϕ yield
|u1 | = ηε v cos (ϕ/2) and |u2 | = ηε v sin (ϕ/2) , (4.3.7)

which give
|u1 |2 + |u2 |2 = ηε2 v 2

1 4 4
|u1 |2 |u2 |2 = η v {1 − cos2 ϕ} (4.3.8)
4 ε
1 4 4
|u1 |4 + |u2 |4 = η v {1 + cos2 ϕ} .
2 ε
Since u1 and u2 are real and do not change sign, we have |∇u1 |2 = |∇|u1 ||2 and
|∇u2 |2 = |∇|u2 ||2 . The relations in (4.3.7) give then
1
|∇u1 |2 + |∇u2 |2 = |∇(vηε )|2 + (vηε )2 |∇ϕ|2 . (4.3.9)
4
Replacing (4.3.8) and (4.3.9) in Eε (u1 , u2 ) we get
Z
1 1
Eε (u1 , u2 ) = |∇(vηε )|2 + 2 V ηε2 v 2 (4.3.10)
2 ε
Z
1 1 2 2 1 1
+ v ηε |∇ϕ|2 + 2 ηε4 v 4 {1 + cos2 ϕ} + gε ηε4 v 4 {1 − cos2 ϕ} .
2 4 4ε 4
The previous formulation of the energy is the one given by the spin formulation
(see the introduction and [126]). We now show how the phase transition model
is obtained. Performing an integration by parts, using (4.2.1) and the first mass
constraint in (4.1.7), we obtain
Z Z
1 1
2
|∇(vηε )| + 2 V ηε2 v 2 = v 2 ηε − ∆ηε + 2 V ηε + ηε2 |∇v|2
ε ε
Z 1 1 1
= v 2 ηε − ∆ηε + 2 V ηε + 2 ηε3 − 2 ηε4 v 2 + ηε2 |∇v|2
ε ε ε
Z Z
λε 1
= 2 v 2 ηε2 + ηε2 |∇v|2 − 2 ηε4 v 2 (4.3.11)
ε ε
Z
λε 1
= 2 + ηε2 |∇v|2 − 2 ηε4 v 2 .
ε ε
Using again (4.2.1), together with the mass constraint for ηε , we have that
Z
λε 1 4

= 2 Eε (ηε ) + 2 ηε . (4.3.12)
ε2 4ε
Replacing (4.3.12) in (4.3.11), and then (4.3.11) in (4.3.10) we get
Z
1 1
Eε (u1 , u2 ) = Eε (ηε ) + ηε2 |∇v|2 + 2 ηε4 {1 − 2v 2 }
2 2ε
Z
1 1 2 2 1 1
+ v ηε |∇ϕ|2 + 2 ηε4 v 4 {1 + cos2 ϕ} + gε ηε4 v 4 {1 − cos2 ϕ} .
2 4 4ε 4
Completing the square for {1 − v 2 } we get
Z
1 1 1
Eε (u1 , u2 ) = Eε (η) + ηε2 |∇v|2 + 2 ηε4 {1 − v 2 }2 − 2 ηε4 v 4
2 2ε 2ε
Z
1 1 2 2 1 1
+ v ηε |∇ϕ|2 + 2 ηε4 v 4 {1 + cos2 ϕ} + gε ηε4 v 4 {1 − cos2 ϕ}
2 4 4ε 4
Z
1 1
= Eε (η) + ηε2 |∇v|2 + 2 ηε4 {1 − v 2 }2
2 2ε
Z
1 1 2 2 1 1
+ v ηε |∇ϕ|2 + ηε4 v 4 gε 1 − {1 − cos2 ϕ} ,
2 4 4 gε ε2
which finishes the proof.

4.4. UPPER BOUND INEQUALITY 105
4.4 Upper bound inequality

In this section, we consider the formulation of the problem in (v, ϕ) and call
Fε (v, ϕ) = Fε (v) + Gε (v, ϕ).
We prove here the upper bound inequality for εFε :
Proposition 4.5. (Upper bound inequality for εFε ) Let ϕ = π1A ∈ X. There
is a sequence of pairs (vε , ϕε ) ∈ Lip(R2 ; (0, 1] × [0, π]), converging as ε → 0 to (1, ϕ)
in L1loc (D) × L1loc (D), such that
lim sup εFε (vε , ϕε ) ≤ F(ϕ) .

ε→0
The proof is based on Bouchitté’s paper [40], where he proves the Γ-convergence of
an anisotropic phase transition Cahn-Hilliard energy. We point out that our weight
ηε depends on ε and vanishes asymptotically on the boundary of D.
In a first step, we assume that ϕ = π1A , where A is an open bounded subset of
R2 with smooth boundary such that H1 (∂A ∩ ∂D) = 0. Then, for any ϕ ∈ X we
approximate {ϕ = π} by this kind of sets. We conclude then thanks to a density
argument. We remark that we do not consider here the mass constraints in (4.1.7).
Before proving the upper bound, we recall some results about sets with smooth
boundary, that can be found in Lemmas 3 and 4 of [130]. For an open set A ⊂ R2
with smooth, non empty compact boundary, let d be the signed distance to ∂A,
defined by

dist(x, ∂A) if x ∈ A
d(x) =
−dist(x, ∂A) if x ∈ R2 \A .
For small t > 0, consider the neighborhood of ∂A given by
Nt = {x ∈ R2 ; |d(x)| < t} ,
with boundary
St = {x ∈ R2 ; |d(x)| = t} .
For t > 0 small enough, there is a diffeomorphism Φ between Nt and ∂A×]0, t[ such
that
∃b > 0 , det |DΦ| ≥ b . (4.4.1)

We denote by Φ̂ the component of Φ in ∂A. Moreover, d is a Lipschitz continuous
function in Nt and we have that
|Dd| = 1 a.e. in Nt . (4.4.2)

For small t > 0, define the measure
µt = H1 x (D ∩ St ) .
Notice that µ0 = H1 x (D ∩ ∂A). As in Lemma 4 in [130], H1 (∂A ∩ ∂D) = 0 yields
lim inf µt (Ω) ≥ µ0 (Ω) ,

t→0
2
for every open Ω ⊂ R , and
lim µt (D) = µ0 (D) . (4.4.3)

t→0
Hence, as t → 0, µt converges weakly∗ to µ0 , which implies

Z Z
lim sup u dµt ≤ u dµ0 (4.4.4)
t→0 D D
for every upper semicontinuous function u : D → R with compact support (see

Propositions 1.62 and 1.80 in [21]).
√
Denote η0 = ρ and for ε ≥ 0 define fε : R2 × R × R2 → R+ by
1 1
fε (x, t, p) = ηε2 (x)|p|2 + ηε4 (x){1 − t2 }2 .
2 4
For |p| = 1 and s ∈ R we also write fε (x, t, s) = fε (x, t, sp).
The last step in the proof of Proposition 4.5 uses the following Lemma, which proof
is given in the appendix.
Lemma 4.5. Let A be a subset of D with 1A ∈ BVloc (D). There exists a sequence
{Ak }k∈N of open bounded subsets of R2 with smooth boundaries such that :
(i) limk→∞ Λ2 ((Ak ∩ D)∆A) = 0,
R R
(ii) lim supk→∞ D ρ3/2 |D1Ak | ≤ D ρ3/2 |D1A |,
R R
(ii) Ak ∩D ρ = A ρ and H1 (∂D ∩ ∂Ak ) = 0 for k large enough.
Proof of Proposition 4.5 : we first assume that A is an open subset of R2 with

smooth, non empty compact boundary such that
H1 (∂A ∩ ∂D) = 0 . (4.4.5)

(Step 1 : construction of the pairs of test functions.) For T > 1, consider the ap-
proximation of the optimal profile

 tanh in (0, T )
wT = h in (T, T + 1/T )
1 in (T + 1/T , +∞).

where h is the unique cubic polynomial such that h(T ) = tanh(T ), h0 (T ) = tanh0 (T ),
h(T + 1/T ) = 1 and h0 (T + 1/T ) = 0. Computing explicitly the coefficients of h, we find
that wT is a nondecreasing function in R+ , with uniform C 1 -bounds with respect to
T ∈ (1, ∞). We extend wT to the whole real line by setting wT (t) = wT (−t) in R− .
For ε ≥ 0, consider 0 < mε ε (to be chosen later), tε = tanh−1 (mε ) and define a
modification of wT near zero by

mε in (0, tε )
wε,T =
wT in (tε , ∞) .
Notice that wε,T has uniform Lipchitz bounds with respect to T ∈ (1, ∞) and
ε ∈ [0, 1). We recall that D = B(0, λ) and we denote Dδ = B(0, λ − δ). For y in
∂A\Dδ , we define
r !
y ρ(y)
wε,T (t) = wε,T t ,
2
and we write wTy = w0,T
y
. For small δ > 0, we define R = Rδ by
r
2
R = (T + 1/T ) . (4.4.6)
δλ
Since wT has uniform C 1 -bounds with respect to T ∈ (1, ∞), while ρ is a smooth
function in Dδ , for every y ∈ ∂A ∩ Dδ , there is an open neighborhood S of y in
∂A ∩ Dδ such that
Z R Z R
f0 (x, wTy (t), (wTy )0 (t)) dt ≤ f0 (x, wTx (t), (wTx )0 (t)) dt + δ ∀x ∈ S , ∀T ≥ 1 .
0 0
Hence, thanks to the compactness of ∂A ∩ Dδ , there is a finite family {Si }N

i=1 of open
disjoint subsets of ∂A ∩ Dδ , and a corresponding family of points yi ∈ Si , such that
N
!
[
H1 ∂A ∩ Dδ \ Si = 0 (4.4.7)
i=1
and
Z R Z R
f0 (x, wTyi (t), (wTyi )0 (t)) dt ≤ f0 (x, wTx (t), (wTx )0 (t)) dt + δ , (4.4.8)
0 0
yi
for every x ∈ Si , T ≥ 1 and 1 ≤ i ≤ N . We will use the functions wε,T to define
the first test function, so we have to interpolate between the different Si ’s. Define
first S0 = ∂A\Dδ and y0 = (λ − δ, 0) ∈ ∂Dδ . For small ` > 0 define Si` = {x ∈
Si ; dist(x, ∂Si ) ≥ `}. Clearly,
H1 (Si \Si` ) → 0 as ` → 0.
In particular, we can take ` = `δ such that
R H1 (Si \Si` ) = oδ→0 (1) (4.4.9)

∞
for every 0 ≤ i ≤ N . Consider then {θ̂i }N
i=0 such that θ̂i ∈ C (∂A, [0, 1]),
N
X
θ̂i = 1 on ∂A and θ̂i = 1 in Si` . (4.4.10)
i=0
We deduce a smooth partition of the unity on Nt by setting θi = θ̂i ◦ Φ̂ and we define



 1 in R2 \NεR
vε = .
 PN θi (x) wyi |d(x)|

in NεR
i=0 ε,T ε
yi
Since wε,T is a nondecreasing function, while ρ is a radial decreasing function, (4.2.8)
and the fact that dist(yi , D) ≥ δ yield
r !
yi
yi ρ(yi )
wε,T ∂NεR
= wε,T (R) = wε,T (T + 1/T ) ≥ wε,T (T + 1/T ) = 1 , (4.4.11)
δλ
so vε is a continuous function. Moreover, since wε,T has uniform Lipschitz bounds

with respect to T ∈ (1, ∞) and ε ∈ (0, 1), there is C > 0 such that for ε small
enough,
C
kvε kC 0,1 (R2 ) ≤ . (4.4.12)
ε
We also define

 π if x ∈ A\Nεt̃ε
ϕε (x) = ξ(d(x)/εtε ) if x ∈ Nεt̃ε , (4.4.13)
0 if x ∈ R2 \(A ∪ Nεt̃ε )

where ξ(t) = π/2(1+t) and t̃ε = (2/λ)1/2 tε . We clearly have that (vε , ϕε ) ∈ Lip(R2 ; (0, 1]×
[0, π]), and that (vε , ϕε ) converges as ε → 0 to (1, ϕ) in L1loc (D) × L1loc (D).
(Step 2 : estimating the energy εGε .) The function ϕε is constant in R2 \Nεt̃ε , so

Gε (vε , ϕε ) = Gε (vε , ϕε ; Nεt̃ε ). Since wε,T is a nondecreasing function while ρ has a
global maximum at zero, for every x ∈ Nεt̃ε we have

yi |d(x)| 0 |d(x)| λ |d(x)|
wε,T ≤ wε,T = wε,T √ ≤ wε,T (tε ) = mε ,
ε ε 2 ε
so vε ≤ mε in Nεt̃ε . Hence,
Z
1
Gε (vε , ϕε ) ≤ ηε2 m2ε |∇ϕε |2 + ηε4 m4ε g̃ε {1 − cos2 (ϕε )} .
8 Nεt̃ε
Then, the definitions of ϕε and g̃ε , together with the fact that ηε is uniformly boun-
ded, yield
Gε (vε , ϕε ) ≤ C m2ε (εt̃ε )−2 + gε m4ε Λ2 (Nεt̃ε ) .

Using (4.4.1), we have
Z εt̃ε Z
−1 −1 det (DΦ)−1 dH1
2

(εt̃ε ) Λ (Nεt̃ε ) ≤ C (εt̃ε ) dt
−εt̃ε ∂A
≤ b−1 H1 (∂A)
= Oε→0 (1) .
For ε small, we have t̃ε = (2/λ)1/2 tanh−1 (mε ) = Oε→0 (mε ). Hence,
Gε (vε , ϕε ) ≤ C mε ε−1 + m5ε gε ε .

−1/4 −3/2
Taking mε = (gε ε2 )−1/4 , Gε (vε , ϕε ) ≤ Cgε ε , and after (4.1.5) we obtain
εGε (vε , ϕε ) ≤ C(gε ε2 )− /4 = oε→0 (1) .

1
(4.4.14)
(Step 3 : computing the energy εFε .) Since vε is constant out of NεR , we have
Z Z Z
εFε (vε ) = φε + φε + φε . (4.4.15)
NεR ∩Dδ NεR ∩(D\Dδ ) NεR \D
where
1
φε (x) = fε (x, vε , εDvε ) .
ε
Considering (4.4.12), for ε small enough there is C > 0 such that |∇vε | ≤ C/ε.
Estimates (4.2.4), (4.2.7) and (4.2.8) thus yield
Z
φε ≤ C max{ρ + ρ } + Cδ ε | ln ε| ε−1 Λ2 (NεR ∩ (D\Dδ ))
2 2
NεR ∩(D\Dδ ) ∂Dδ
≤ C δ + ε2 | ln ε| ε−1 Λ2 (NεR ) .

Using (4.4.1) we have
Z εR Z
−1 −1 det (DΦ)−1 dH1
2

ε Λ (NεR ) ≤ C ε dt
−εR ∂A
−1 1
≤ C Rb H (∂A) ,
so (4.4.6) yields
Z
lim φε ≤ C δ R b−1 H1 (∂A) = oδ→0 (1) . (4.4.16)
ε→0 NεR ∩(D\Dδ )
Similarly, using (4.2.6) we have
Z
φε ≤ C sup ηε2 + ηε4 R b−1 H1 (∂A)

NεR \D R2 \D
≤ C ε /3 R b−1 H1 (∂A)
1
(4.4.17)
= oε→0 (1) .
Now, remember the interpolation from (4.4.10). We have

Z N Z
X Z
φε = φε + φε ,
NεR ∩Dδ i=1 NεR ∩Bi NεR ∩Ci
where
Bi = {x ∈ Dδ ; Φ̂(x) ∈ Si` } and Ci = {x ∈ Dδ ; Φ̂(x) ∈ Σi \Si` } .

As before, since ηε is uniformly bounded in R2 with respect to ε ∈ (0, 1), (4.4.1)
yields
Z Z εR Z
1 det (DΦ)−1 dH1

φε ≤ C dt
NεR ∩Ci ε −εR Σi \Si`
≤ C R b−1 H1 (Σi \Si` ) .
Hence, after (4.4.6) and (4.4.9), we obtain

Z
φε = oδ→0 (1) (4.4.18)
NεR ∩Ci
yi |d(x)|
for every ε ∈ (0, 1). In NεR ∩ Bi we have vε (x) = wε,T ( /ε). Using (4.4.2) we write
Z Z
yi |d(x)| yi 0 |d(x)|
φε = |D(|d|/ε)|(x)fε (x, wε,T ( /ε), (wε,T )( /ε)) .
NεR ∩Bi NεR ∩Bi
The coarea formula from Proposition 4.1 yields

Z Z R Z
yi yi 0
φε = dt fε (x, wε,T (t), (wε,T ) (t)) dµεt (x) .
NεR ∩Bi −R Dδ ∩Bi
We thus have,
Z Z R Z
φε ≤ dt f0 (x, wTyi (t), (wTyi )0 (t)) dµεt (x) + Rεi + R̃εi . (4.4.19)
NεR ∩Bi −R Dδ ∩Bi
The first error here before comes from the modification of wT near 0. Using (4.4.3)
and the definition of tε we compute
s
2
Rεi ≤ C tε sup kµt k
ρ(yi ) t∈(0,εR)
r
2
≤ C tε sup kµt k
δλ t∈(0,εR)
= oε→0 (1) .
The second error appears when replacing fε by f0 , so using estimates (4.2.4) and
(4.2.8), together with yi ∈ Dδ , there is Cδ > 0 such that
Rε2 ≤ Cδ ε2 | ln ε| R supt∈(0,εR) kµt k = oε→0 (1) .
Using Fubini’s formula, we rewrite (4.4.19) as
Z Z Z R
φε ≤ 1Dδ ∩Bi (x) f0 (x, wTyi (t), (wTyi )0 (t)) dt dµεt (x) + oε→0 (1) .
NεR ∩Bi D −R
The set Dδ ∩Bi is close and the inner integral is a continuous function of x. Hence, the
function inside the outer integral is upper semicontinuous function of x. Inequality
(4.4.4) thus yields
Z Z Z R
yi yi 0
lim sup φε ≤ 1Dδ ∩Bi (x) f0 (x, wT (t), (wT ) (t)) dt dµ0 (x)
ε→0 NεR ∩Bi D −R
Z Z R
yi yi 0
= f0 (x, wT (t), (wT ) (t)) dt dµ0 (x) .
Dδ ∩Si` −R
Notice that since µ0 is supported in ∂A, we replaced Bi by Dδ ∩ Si` . After (4.4.8)

and since wT is an even function, we have
Z Z Z R
lim sup φε ≤ 2 f0 (x, wTx (t), (wTx )0 (t)) dt
+ δ dµ0 (x)
ε→0 NεR ∩Bi Dδ ∩Si` 0
Z Z ∞
x x 0
≤ 2 f0 (x, wT (t), (wT ) (t)) dt + δ dµ0 (x) .
(4.4.20)
Dδ ∩Si` 0
(Step 4 : upper bound) Putting together (4.4.14)-(4.4.18) and (4.4.20) we get
N Z
X Z ∞
lim sup εFε (vε , ϕε ) ≤ 2 f0 (x, wTx (t), (wTx )0 (t)) dt + δ dµ0 (x) + oδ→0 (1) .
ε→0
i=1 Dδ ∩Si` 0
Now, we take a sequence T = Tδ such that Tδ → ∞ as δ → 0 (notice that (4.4.11)

still holds). Then, Fubini’s formula and dominated convergence theorem, together
with (4.4.7) and (4.4.9), yield
Z Z ∞
x x 0
lim lim sup εFε (vε , ϕε ) ≤ f0 (x, w (t), (w ) (t)) dt dµ0 (x) .
δ→0 ε→0 D 0
Remembering the definitions of f0 , µ0 and wx , (4.1.13) yields
Z
ρ /2 (x)dH1 (x) .
3
lim lim sup εFε (vε , ϕε ) ≤ 2σ
δ→0 ε→0 D∩∂A
We conclude thanks to a diagonal argument (see Corollary 1.16 in [25]) : there exists
a sequence δε →ε→0 0, such that as ε → 0, (vε,δε , ϕε,δε ) converges in L1loc (D)×L1loc (D)
to (1, ϕ), and
Z
ρ /2 (x)dH1 (x) .
3
lim sup εFε,δε (vε,δε , ϕε,δε ) ≤ 2σ
ε→0 D∩∂A
(Step 5 : approximation of A by Cacciopoli sets) We end the proof using Lemma 4.5,
the proof of which is given in the appendix. We remove the condition (4.4.5) and we
only assume that A is a set with locally finite perimeter in D. Consider ϕk = π1Ak
with {Ak }k∈N the sequence

R from Lemma 4.5. After (i), (1, ϕk ) converges to (1, ϕ) in
L1 (D) and we have Ak ρ = α2 for k large enough. Hence, from steps (1)-(4), there
is a sequence (vεk , ϕkε ) → (1, ϕk ) in L1loc (D) × L1loc (D) such that
lim εFε (vεk , ϕkε ) = F(ϕk ) .

ε→0
Using (ii) from Lemma 4.5 we obtain

k k
lim sup lim εFε (vε , ϕε ) ≤ F(ϕ) .
k→0 ε→0
As in step (4), we conclude thanks to a diagonal argument.
4.5 Lower bound inequality and compactness

The proofs in this section are based on geometric measure theory techniques. We
make the lower bound on lines and then use the slicing method, which can be found
in [22] or [42]. The last part of the proof of the lower bound is inspired by the ideas
in [18].
4.5.1 Lower bound on lines

Consider an open set A ⊂⊂ D and let ν ∈ S 1 be a fixed direction. We call πν the
hyperplane orthogonal to ν, and Aν the projection of A on πν . We define the one
dimensional slices of A, indexed by x ∈ Aν , as
Ax = {t ∈ R ; x + tν ∈ A} .
For every function f in D, we define fx as the restriction of f to the slice Ax , defined
by fx (t) = f (x + tν). For (v, ϕ) : Ax → (0, 1] × (0, π), we define the energies
Z
1 2 1 4
Fε (v ; Ax ) = ηε,x |∇v|2 + 2 ηε,x {1 − v 2 }2 ,
2 Ax 2ε
Z
1 2
Gε (v, ϕ ; Ax ) = ηε,x v 2 |∇ϕ|2 + ηε,x
4
v 4 g̃ε {1 − cos2 (ϕ)} and
8 Ax
Fε (v, ϕ ; Ax ) = Fε (v; Ax ) + Gε (v, ϕ; Ax ) .
Similarly, for ϕ ∈ BV (Ax ; {0, π}) we define

Z
2σ
ρ /2 d|Dϕ| .
3
F(ϕ ; Ax ) =
π Ax x
With the previous notations we have the following result :
Proposition 4.6. Let (vε , ϕε ) ∈ Lip(Ax ; (0, +∞) × [0, π]) such that
sup kvε kL∞ (Ax ) < ∞ (4.5.1)

ε>0
4.5. LOWER BOUND INEQUALITY AND COMPACTNESS 113
and
εFε (vε , ϕε ; Ax ) < ∞ . (4.5.2)

Then, there is ϕ ∈ SBV (Ax ; {0, π}) such that
(vε , ϕε ) → (1, ϕ) in L1 (Ax ) × L1 (Ax ) (4.5.3)
and
lim inf εFε (vε , ϕε ; Ax ) ≥ F(ϕ ; Ax ) . (4.5.4)

ε→0
Proof : (Step 1) Using that A ⊂⊂ D and estimate (4.2.4), there are c1 , c2 > 0 such
that
2
ηε,x > ρx − c1 ε2 | ln ε| > c2 in Ax . (4.5.5)
Hence, the definition of Fε (· ; Ax ) and (4.5.2) give
4|Ax | 2
Z
|1 − vε | < ε Fε (vε , ϕε ; Ax ) = oε→0 (1) ,
Ax c22
so vε → 1 in L1 (Ax ). Similarly, after (4.5.1) vε < C in Ax , so (4.1.5) yields
Z
8C
vε4 |1 − cos2 (ϕε )| < Fε (vε , ϕε ; Ax ) = oε→0 (1) .
Ax g̃ε c22
Hence, up to a (not relabeled) subsequence, ϕε → ϕ a.e. in Ax , with ϕ : Ax → {0, π}.
This, together with Ax ⊂⊂ D, gives ϕε → ϕ in L1 (Ax ). We have proved (4.5.3).
(Step 2) We now prove the lower bound for the energy. Let t0 ∈ Sϕ. For δ > 0 define
Jδ = Ax ∩ (t0 − δ, t0 + δ) ,
and suppose that
n o
inf inf vε (t) > c3 > 0 . (4.5.6)
t∈Jδ ε>0
Then, for every ε > 0 and every t ∈ Jδ , vε (t) > c3 . Hence, using (4.5.2), (4.5.5) and
the coarea formula (4.1.20), there is C > 0 such that
Z Z π Z
8C
(c−1 −2 −2 −4 2 2 2 1/2
2 c3 +c2 c3 ) ≥ |∇ϕ| +{1−cos (ϕ)} ≥ dt {1−cos (t)} |D1Wε,t | ,
g̃ε ε2 Jδ 0 Jδ
(4.5.7)
where Wε,t = {t ∈ Ax ; ϕε (x) < t}. Since ϕε converges to ϕ a.e. in Ax , we get
1Wε,t → 1{ϕ=0} in L1 (Ax ) ,

for a.e. t ∈ (0, π). Hence, the lower semicontinuity of the BV norm with respect to
the L1 -convergence, together with (4.1.5), (4.5.7) and Fatou lemma, gives
Z π Z
2 1/2
0 ≥ dt {1 − cos (t)} lim inf |D1Wε,t |
0 ε→0 Jδ
Z
π
≥ |D1{ϕ=0} | .
2 Jδ
Thus,
0 = H0 (Jδ ∩ Sϕ) ≥ H0 ({t0 }) = 1 .

This contradiction implies that (4.5.6) can not be satisfied. We derive that for every
δ > 0, we may extract a subsequence (not relabeled), such that exists {tε }ε>0 ⊂ Jδ
with
tε → t̃0 ∈ Jδ and vε (tε ) → 0 as ε → 0 . (4.5.8)

(Step 3) For ε > 0, define Iε± = {t ∈ Jδ ; ±(tε − t) < 0} and vε± : Jδ → (0, 1] by
vε± (t) = 1Iε± vε (tε ) + 1Iε∓ vε (t) .

The definition of Fε , estimate (4.5.5) and the fact that vε± is constant in Iε± while
equal to vε in Iε∓ yield
√ Z
ρx/2 |(vε+ )0 | |1 − (vε+ )2 | + |(vε− )0 | |1 − (vε− )2 | + oε→0 (1) .
3
2 εFε (vε , ϕε ; Jδ ) ≥
Jδ
Using the coarea formula (4.1.20) we obtain
√ Z 1
2
Z
ρx/2 |D1Vε,t
3
2 εFε (vε , ϕε ; Jδ ) ≥ dt (1 − t ) + | + |D1 − |
Vε,t + oε→0 (1) ,
0 Jδ
±
where Vε,t = {t ∈ Jδ ; vε± < t}. Since tε → t̃0 , vε (tε ) → 0 and vε (t) → 1 a.e. in
1 ±
± → 1I ∓ in L (Jδ ), where I
Jδ , 1Vε,t = {t ∈ Jδ ; ±(t̃0 − t) ≤ 0}. Hence, the lower
semicontinuity of the BV norm with respect to the L1 -convergence and Fatou lemma
give
Z 1 Z
1 2 3/2

|D1I − | + |D1I + | = σ 2ρx/2 (t̃0 ) .
3
lim inf εFε (vε , ϕε ; Jδ ) ≥ √ dt (1 − t ) ρx
ε→0 2 0 J
(4.5.9)
3/2
Moreover, since ρx ≥ c4 > 0 in Ax , we have
lim inf εFε (vε , ϕε ; Jδ ) ≥ 2σc4 > 0 . (4.5.10)

ε→0
(Step 4) Let Γ = {t1 , . . . , tn }, n ∈ N, be any finite subset of Sϕ. For i ∈ {0, 1, . . . , n}

we define
Jδi = Ax ∩ (ti − δ, ti + δ) .
Consider δ 0 > 0 such that Jδi ∩ Jδj = ∅ for i 6= j and let δ ∈ (0, δ 0 ). After (4.5.10), we
have
c4
lim inf εFε (vε , ϕε ; Ax ) ≥ n .
2σ ε→0
Therefore, using (4.5.2) we derive that n is bounded, so Sϕ is a finite set and
ϕ ∈ SBV (Ax ).
(Step 5) Finally, write Sϕ = {t1 , · · · , tN }, N ∈ N. Reasoning as before, for δ 0 small

enough and δ ∈ (0, δ 0 ), (4.5.9) gives
N
X
ρx/2 (t̃i ) ,
3
lim inf εFε (vε , ϕε ; Ax ) ≥ 2σ with t̃i ∈ Jδi .
ε→0
i=1
Since ρx is a continuous function, taking the limit δ → 0 in the previous inequality

we obtain
N
X Z
3/2
ρx/2 d|Dϕ| .
3
lim inf εFε (vε , ϕε ; Ax ) ≥ 2σ ρx (ti ) = 2σ
ε→0 Ax
i=1
We have proved (4.5.4).
4.5.2 The slicing method

Using the slicing method, we now prove the compactness and the lower bound in-
equality for εFε .
Proposition 4.7. (Lower bound inequality and compactness for εFε ) Let
(vε , ϕε ) ∈ Liploc (R2 ; (0, +∞) × [0, π]) such that
sup kvε kL∞ (K) < CK for K ⊂⊂ D (4.5.11)

ε>0
and
sup εFε (vε , ϕε ) < ∞ . (4.5.12)
ε>0
Then, there is ϕ ∈ X such that
(vε , ϕε ) → (1, ϕ) in L1loc (D) × L1loc (D) (4.5.13)
and
lim inf εFε (vε , ϕε ) ≥ F(ϕ) . (4.5.14)

ε→0
Proof : arguing as in the proof of (4.5.3) of Proposition 4.6, there exists ϕ : D →

{0, π} such that (4.5.13) is satisfied. Consider an open set A ⊂⊂ D, and fix ν ∈ S 1 .
For x ∈ Aν , we define (vε,x , ϕε,x ) : Ax → (0, 1] × (0, π) by
(vε,x , ϕε,x )(t) = (vε , ϕε )(x + tν) .

Since A ⊂⊂ D and since vε , ηε are non vanishing continuous functions, for fixed
ε > 0 (4.5.12) yields
Z Z
ε 2 2 ε 2 2
C ≥ εFε (vε , ϕε ; A) ≥ inf η |∇vε | + inf vε ηε |∇ϕε |2
2 A ε A 8 A
Z Z A
≥ cε,A |∇vε |2 + |∇ϕε |2 ,
A A
so vε and ϕε belong to W 1,2 (A). Hence see [74], section 4.9.2),
0
vε,x (t) = Dν vε (x + tν) and ϕ0ε,x (t) = Dν ϕε (x + tν)
for a.e. t ∈ Ax , for Λ1 - a.e. x ∈ Aν . Using then |∇vε |2 ≥ |Dν vε |2 , we get the slicing
inequality
Z
εFε (vε , ϕε ; A) ≥ εFε (vε,x , ϕε,x ; Ax ) dx . (4.5.15)
Aν
After (4.5.12), for Λ1 -a.e. x ∈ Aν , εFε (vε,x , ϕε,x ; Ax ) is uniformly bounded with res-
pect to ε. Thus, after Proposition 4.6, for Λ1 -a.e. x ∈ Aν there is ϕx ∈ BV (Ax ; {0, π})
such that
(vε,x , ϕε,x ) → (1, ϕx ) in L1 (Ax ) × L1 (Ax ) (4.5.16)
and
lim inf εFε (vεx , ϕε,x ; Ax ) > F(ϕx ; Ax ) . (4.5.17)

ε>0
The function ϕ defined in (4.5.13) is the L1 (A) limit of ϕε , so for Λ1 -a.e. x ∈

Aν , ϕx coincide with the restriction of ϕ to Ax . Therefore, since the vector ν is
taken arbitrarily, ϕ ∈ BV (A) (see Proposition 6.9 in [18]), and since A is any open
relatively compact subset of D, we derive that ϕ ∈ BVloc (D). Using (4.5.15), (4.5.17)
Fatou’s lemma and Fubini’s formula, we also obtain
Z
lim inf εFε (vε , ϕε ) ≥ F(1, ϕx ; Ax ) dx
ε→0 Aν
Z Z
2σ
ρx/2 d|Dϕ|
3
= dx
π Aν Ax
Z
2σ
ρ /2 d(Λ1 x Aν ⊗ |Dϕx | x Ax ) .
3
= (4.5.18)
π D x
Now, for every ε > 0, let µε be the energy distribution in D associated with the pair
(vε , ϕε ), that is, the positive Radon measure which for every Borel set E ⊂ R2 is
given by
µε (E) = εFε (vε , ϕε ; E ∩ D) .

After (4.5.12), the total mass kµε k is uniformly bounded. De La Vallée Poussin
compactness criterion gives then that (up to a subsequence) µε converges weakly∗
to some finite measure µ on D. We claim that
µ ≥ 2σρ /2 · H1 x Sϕ .
3
We will prove this using Besicovitch derivation Theorem. First, after (4.5.12) for
every K ⊂⊂ D there is RK ∈ (0, λ) such that
0 ≤ µ(K) ≤ µ(B(0, RK )) ≤ lim inf µε (B(0, RK )) < ∞. (4.5.19)

ε→0
Hence, µ is a positive Radon measure in D, and for H1 -a.e. x ∈ Sϕ the limit
µ(Br (x))
f (x) = lim+ 1
(4.5.20)
r→0 H (Br (x) ∩ Sϕ)
exists, and we have
µ ≥ f · H1 x Sϕ . (4.5.21)
Let x0 ∈ Sϕ ∩ A. Since A ⊂⊂ D, Br (x0 ) ⊂ D for r small enough. We assume 4
moreover that µ(∂B(x0 , r)) = 0. Proposition 1.62 in [21] and estimate (4.5.18) yield
µ(Br (x0 )) = lim µε (Br (x0 ))

ε→0
Z
2σ
ρ /2 d(Λ1 x Aν ⊗ |Dϕx | x Ax )
3
≥
π Br (x0 ) x
Z
2σ
d(Λ1 x Aν ⊗ |Dϕx | x Ax ) .
3/2
≥ inf ρx (4.5.22)
π Br (x0 ) Br (x)
In Proposition 4.6 we proved that for Λ1 -a.e. x ∈ Aν , ϕx ∈ SBV (Ax ) ∩ L∞ (A), and
Z Z
dx H0 (Sϕx ) < ∞ .
Aν Ax
Hence, after Theorem 2.3 in [22], ϕ ∈ SBV (A) ∩ L∞ (A) and

Z Z
d(Λ x Aν ⊗ |Dϕx | x Ax ) = π
1
|hνϕ , νi| dH1 , (4.5.23)
Br (x0 ) Br (x0 )
where νϕ is the measure theoretic inner normal to the Caccioppoli set {ϕ = π}.
Putting (4.5.23) in (4.5.22) we obtain
Z
3/2
µ(Br (x0 )) ≥ 2σ inf ρx |hνϕ , νi| dH1
Br (x0 ) Br (x0 )
3/2
≥ 2σ inf ρx inf |hνϕ , νi| H1 (Br (x0 ) ∩ Sϕ) .
Br (x0 ) Br (x0 )∩Sϕ
4. In fact this holds for all r except countably many (see [21], page 29).
Since Sϕ is a rectifiable set in A, for H1 -a.e. x0 ∈ Sϕ, νϕ (x) is continuous in Br (x0 )

for r small enough. Thus, taking ν = νϕ (x0 ) and since ρ is continuous, we get
µ(Br (x0 ))
≥ 2σρ /2 (x0 )
3
lim+ 1
r→0 H (Br (x0 ) ∩ Sϕ)
for H1 -a.e. x0 ∈ Sϕ. Hence, (4.5.20) and (4.5.21) yield the claim. The definition of
µ gives then
Z
ρ /2 dH1 .
3
lim inf εFε (vε , ϕε ; A) = lim inf µε (A) ≥ µ(A) ≥ 2σ
ε→0 ε→0 Sϕ∩A
Finally, taking an increasing sequence {Ak }k∈N with Ak ⊂⊂ D, we get

Z
ρ /2 dH1 ,
3
lim inf εFε (vε , ϕε ) ≥ 2σ
ε→0 Sϕ∩D
which gives (4.5.14).
4.5.3 Remark about a lower bound for vε in the transition

zone
We end this section with a discussion about the infinimum of vε in the tran-
sition zone. Let {(vε , ϕε )}ε>0 be a sequence of minimizers of Fε , and let ϕ ∈
BVloc (D ; {0, π}) be the L1loc -limit of ϕε given in (4.1.16). Let K ⊂⊂ D be an open
smooth set, with non negligible intersection with Sϕ, that is,
H1 (K ∩ Sϕ) > 0 .
For every ε > 0, we define
mε,K = inf vε (x) .

x∈K
We would like to obtain an upper bound for mε,K , in connection with an open
question in [31], namely
mε,K ≤ CK (gε ε2 )− /4 .
1
(4.5.24)
If we assume that we have the upper and lower inequalities for each ε > 0, that is
εFε (ṽε ) ≤ F(ϕ) (4.5.25)

and
εFε (ṽε ) ≥ F(ϕ) , (4.5.26)

we can give estimates on Gε in order to obtain the upper bound for mε,K . So assume
that we have (4.5.25) and (4.5.26). On the one hand, estimates (4.2.4) and (4.2.8)
give then
1p 2 3 Z
3/2 2 2
εGε (vε , ϕε ; K) ≥ gε ε mε,K inf ρ − CK ε | ln ε | |∇ϕε | sin ϕε
4 K K
p Z
2 3
≥ CK gε ε mε,K |∇ϕε | sin ϕε .
K
We claim that the integral here below is bounded away from zero. Indeed, if this
not the case, we will have
Z
lim inf |∇ϕε | sin ϕε = 0 .
ε→0 K
1
Hence, since ϕε → ϕ in L (K), the coarea formula together with the lower semi
continuity of the BV norm imply the contradiction
Z π Z Z π Z
0 = lim inf sin t dt |D1{ϕε <t} | ≥ sin t dt |D1{ϕ=0} | = 2 H1 (Sϕ ∩ K) .
ε→0 0 K 0 K
0
We thus derive that there is CK > 0 such that
p
0
εGε (vε , ϕε ; K) ≥ CK gε ε2 m3ε,K . (4.5.27)
In the other hand, by inspection of the proof of Proposition 4.5 (see estimate
(4.4.14)), we see that the pair of test function (ṽε , ϕ̃ε ) satisfies
εGε (ṽε , ϕ̃ε ; K) ≤ C(gε ε2 )− /4 .

1
(4.5.28)
Hence, considering (4.5.25)-(4.5.28), together with the fact that (vε , ϕε ) minimizes
Fε , we obtain
Z p
ρ /2 |Dϕ| + CK gε ε2 m3ε,K ≤ εFε (vε , ϕε ; K)
3
2σ
K
≤ εFε (ṽε , ϕ̃ε ; K)
Z
ρ /2 |Dϕ| + C(gε ε2 )− /4 .
3 1
≤ 2σ
K
Multiplying both sides of the previous inequality by (gε ε2 )1/4 we find the upper bound
(4.5.24) for m3ε,K .
However, we are not able to prove (4.5.25) and (4.5.26) as such because of the error
terms. Indeed, the proof of the upper bound of Theorem 4.1 says that there is a
sequence {(ṽε , ϕ̃ε )}ε>0 such that
lim sup εFε (ṽε ) ≤ F(ϕ) . (4.5.29)

ε→0
In the proof of (4.5.29), we first approximate the locally Cacciopoli set A = {ϕ = π}

by characteristics functions of open sets Ak with compact smooth boundary. This
gives a small error in terms of k ∈ N in the upper bound inequality (4.5.29). Then
for each k ∈ N, we construct a test function for which (4.5.29) holds, up to a small
error term depending on a parameter δ > 0. In these two steps, we use diagonal
extraction arguments in order to get rid of the error terms, so it is not possible to
compute them explicitly. Similarly, in the proof of the lower bound of Theorem 4.1,
we use the compactness of bounded Radon measures, so we cannot estimate the
error term in the lower bound inequality
lim inf εFε (ṽε ) ≥ F(ϕ) . (4.5.30)

ε→0
4.6 Proof of the Γ-convergence for ε (Eε(·) − Eε(ηε))

4.6.1 Proof of the compactness and the lower bound inequa-
lity in Theorem 4.1 :
Let {(u1,ε , u2,ε )}ε>0 be a sequence of minimizers of Eε in H satisfying (4.1.15).
After Proposition 4.3(i), the pairs (vε , ϕε ) are well defined by (4.1.6), belong to
Liploc (R2 ; (0, +∞) × [0, π]) and satisfy (4.5.11). Proposition 4.4 yields εFε (vε , ϕε ) <
∞. Thus, the hypotheses of Proposition 4.7 are fulfilled by (vε , ϕε ) and we have
(vε , ϕε ) → (1, ϕ) in L1loc (D) × L1loc (D)
with ϕ ∈ X, and
lim inf εFε (vε , ϕε ) ≥ F(ϕ ) .

ε→0
Equality (4.1.8) yields then
lim inf ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) ≥ F(ϕ) .

ε→0
Finally, using identity (4.3.4) we get

√
(u1,ε , u2,ε ) → ρ (1{ϕ=0} , 1{ϕ=π} ) in L1loc (D) × L1loc (D) .
In order to prove the upper bound we have to work a little more. We first modify the
pairs of test functions from Proposition 4.5 to make then satisfy the mass constraints
(4.1.7). We prove then that this modification do not change the limit of the energy.
We finish by verifying that the pairs of modified test functions are the image by
(4.1.6) of a pair in H, and we conclude using Proposition 4.5.
4.6.2 Proof of the upper bound inequality in Theorem 4.1 :

(Step 1 : Modification of the pair of test functions) With the notations from the
proof of Proposition 4.5, we write Nε = NεRδ and we define (v̌ε , ϕε ) the sequence of
pairs of test functions such that
lim sup εFε (v̌ε , ϕε ) ≤ F(ϕ) . (4.6.1)

ε→0
4.6. PROOF OF THE Γ-CONVERGENCE FOR ε (Eε (·) − Eε (ηε )) 121
Consider κ ∈ C ∞ (R+ ; [0, 1]) with supp κ ⊂ (0, 1) and κ = 1 in (0, 1/2). Since A is a
non empty open set, there is B0 = Br0 (x0 ) ⊂⊂ A∩D. For ` ∈ [−1, 1] and τ ∈ (1/2, 1),
define κε = κε,`,τ by
κε (x) = ετ ` κ(|x−x0 |/r0 ) .

We define then v̂ε = v̌ε + κε and vε = cε v̂ε , with cε = kηε v̂ε k−2
2 . For ε small enough
Nε and B0 are disjoints. We estimate
Z
c−1
ε = 1+ ηε2 (v̂ε2 − 1)
Nε ∪B0
Z Z Z
2 2 2
= 1+2 ηε κε + ηε (v̌ε − 1) + ηε2 κ2ε
ZB0 Nε B0
= 1+2 ηε2 κε + O(ε) + O(ε2τ ) .

B0
Hence, using that τ ∈ (1/2, 1) we get c2ε = 1 − rε with

Z
rε = 4 ηε2 κε + O(ε) = O(ετ ) . (4.6.2)
B0
Notice that for ε small enough, rε may be positive or negative depending on the sign
of `.
y
The definition of wε,T assures that vε > 0. The first mass constraint in (4.1.7)
is immediately satisfied by the definition of vε . Remember the definition of ϕε in
(4.4.13). For the second mass constraint we write
Z Z
c−2
ε ηε2 vε2 cos ϕε = ηε2 (1R2 \(A∪Nε ) − 1A\(Nε ∪B0 ) + 1Nε ∪B0 v̂ε2 cos ϕε ) .
R2 R2
Adding and removing 1Nε \A ηε2 , 1Nε ∪A ηε2 and 1B0 ηε2 in the previous integral, we get
Z Z Z
c−2
ε ηε2 vε2 cos ϕε = ηε2 (1R2 − 21A ) + ηε2 (v̌ε + κε )2 − 1 (4.6.3)
R2 R2 B0
Z
+ ηε2 (v̌ε2 cos ϕε − 1A + 1R2 \A ) . (4.6.4)
Nε
For the third term in (4.6.3), we have that ηε , v̌ε and cos ϕε are bounded while
L2 (Nε ) = O(ε). Hence,
Z
ηε2 (v̌ε2 cos ϕε − 1A + 1R2 \A ) = O(ε) . (4.6.5)
Nε
R R
For the first term in (4.6.3), using that R2
ηε2 = 1 = α1 + α2 and that D∩A
ρ = α2 ,
we obtain
Z Z Z
ηε2 (1R2 − 21A ) = α1 − α2 + (ηε2 − ρ) + ηε2 .
R2 A∩D A\D
Using (4.2.5) we get, for α ∈ (1/2, 3/5) and γ ∈ (1/2, 3/4),
Z Z Z
(ηε2 − ρ) = (ηε2 − ρ) + (ηε2 − ρ)
A∩D A∩B(0,λ−εα ) (A∩D)\B(0,λ−εα )
= O(εγ ) + O(εα ) . (4.6.6)
Moreover, from (4.2.7), we have ηε2 (x) ≤ ηε2 (xα ) in A\D, with xα ∈ ∂B(0, λ − εα ).
From (4.2.5) and (4.2.8) we get
ηε2 (x) ≤ ηε2 (xα ) = ηε2 (xα ) − ρ(xα ) + ρ(xα ) = O(εα ) ,

so using that A is a bounded set we obtain
Z
ηε2 = O(εα ) . (4.6.7)
A\D
For the second term in (4.6.3), the definitions of κε and rε yield
Z
1
ηε2 κε (2 + κε ) = rε + O(ε) + O(ε2τ ) . (4.6.8)
B0 2
Putting (4.6.5)-(4.6.8) in (4.6.3) and considering (4.6.2) we get

Z
−2 1
cε ηε2 vε2 cos ϕε = α1 − α2 + rε + O(εβ ) ,
R2 2
where β = min{1, α, γ, 2τ } = min{α, γ} ∈ (1/2, 3/5). Hence, (4.6.2) gives
Z
2 2 1
ηε vε cos ϕε − (α1 − α2 ) = − (α1 − α2 ) rε + O(εβ ) .
R2 2
Suppose now, without loss of generality, that α1 − α2 ≤ 1/2. The definition of rε and
κε , together with (4.2.4), (4.2.8) and B0 ⊂⊂ A ∩ D, give then
Z
|rε | ≥ 4 inf ηε2 κ2ε + O(ε)
B0 B0
Z
2 τ
≥ 4 inf ηε |`| ε κ2ε + O(ε)
B0 Br0/2 (x0 )
τ
≥ c |`| ε + O(ε) ,
for some c > 0 not depending on ε. Hence, if we take ` = 1 in the definition of κε ,

for ε small enough we have
Z
ηε2 vε2 cos ϕε − (α1 − α2 ) ≥ c0 ετ (1 + ε1−τ − εβ−τ ) .
R2
Analogously, taking now ` = −1, we get

Z
00
ηε2 vε2 cos ϕε − (α1 − α2 ) ≤ c ετ (−1 + ε1−τ + εβ−τ ) .
R2
Since β ∈ (1/2, 3/5), we can choose τ ∈ (1/2, β) and obtain
Z
ηε2 vε2 cos ϕε > α1 − α2 if ` = 1
R2
and
Z
ηε2 vε2 cos ϕε < α1 − α2 if ` = −1 .
R2
Hence, there exists `ε ∈ (−1, 1) such that for ε small enough, the associated pair
(vε , ϕε ) satisfy the second mass constraint in (4.1.7).
(Step 2 : Computing the energy). We now compute the energy of (vε , ϕε ). We recall
that Nεt̃ε is the transition zone of ϕε defined in (4.4.13). For the energy Gε , we have
that ϕε is constant out of Nεt̃ε , while vε = cε v̌ε in Nεt̃ε with cε = 1 + O(ετ ). Hence,
εGε (vε , ϕε ) = (1 + O(ετ ))εGε (v̌ε , ϕε ) . (4.6.9)

For the energy Fε , we have that vε = cε (1 + κε ) in B0 . The definition of κε gives
then, |∇vε |2 = O(ε2τ ) and {1 − vε2 }2 = O(ε2τ ). Hence,
εFε (vε ; B0 ) = O(ε2τ −1 ) = oε→0 (1) . (4.6.10)

2
In R \(Nε ∩ B0 ), we have that vε = cε , so |∇vε | = 0 and {1 − vε2 }2 = O(ε2τ ). As
before we get
εFε (vε ; R2 \(Nε ∩ B0 )) = O(ε2τ −1 ) = oε→0 (1) . (4.6.11)

In Nε , we have that vε = cε v̌ε . Hence, |∇vε |2 = (1 + O(ετ ))|∇v̌ε |2 and {1 − vε2 }2 =
(1 + O(ετ )){1 − v̌ε2 }2 + O(ετ ), which gives
εFε (vε ; Nε ) = (1 + O(ετ )) εFε (v̌ε ; Nε ) + O(ετ )ε−1 Λ2 (Nε )

= (1 + O(ετ )) εFε (v̌ε ; Nε ) + oε→0 (1) . (4.6.12)
Since v̌ε is constant out of Nε , we have Fε (v̌ε ) = Fε (v̌ε ; Nε ). Putting together (4.6.1)
and (4.6.9)-(4.6.12), we obtain
lim sup εFε (vε , ϕε ) = lim sup εFε (v̌ε , ϕε ) ≤ F(ϕ) . (4.6.13)
ε→0 ε→0
(Step 3 : identification of (vε , ϕε )) The pairs of test functions satisfies the hypothesis
from Proposition 4.3(ii), so defining (u1,ε , u2,ε ) by (4.3.4) we have (u1,ε , u2,ε ) ∈ H
and u21,ε + u22,ε > 0. Hence, after Proposition (4.4) relation (4.1.8) holds, and (4.6.13)
yield
lim sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) = lim sup εFε (vε , ϕε ) ≤ F(ϕ) .
ε→0 ε→0
4.6.3 Proof of Corollary 4.2 :

let ϕ̃ ∈ X with F(ϕ̃) < +∞. After the upper bound inequality in Theorem 4.1,
there is a sequence (ũ1,ε , ũ2,ε ) ∈ H such that
lim sup ε (Eε (ũ1,ε , ũ2,ε ) − Eε (ηε )) ≤ F(ϕ̃) .

ε→0
Since (u1,ε , u2,ε ) minimize Eε in H, the previous inequality yields
lim sup ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) ≤ F(ϕ̃) , (4.6.14)

ε→0
so in particular (u1,ε , u2,ε ) satisfy (4.1.15). Hence, from the compactness and the
lower bound inequality in Theorem 4.1, there is ϕ ∈ X and a subsequence (u1,ε0 , u2,ε0 )
with
lim0 inf ε0 (Eε0 (u1,ε0 , u2,ε0 ) − Eε0 (ηε0 )) ≥ F(ϕ) .

ε →0
This inequality is verified for every subsequence of (u1,ε , u2,ε ), so we have

ε→0
From (4.6.14) and (4.6.15) we obtain
F(ϕ̃) ≥ lim sup εEε (u1,ε , u2,ε ) ≥ lim inf εEε (u1,ε , u2,ε ) ≥ F(ϕ) , (4.6.16)
ε→0 ε→0
so F(ϕ) = inf X F. Taking ϕ̃ = ϕ in (4.6.16), yields
lim ε (Eε (u1,ε , u2,ε ) − Eε (ηε )) = inf F .

ε→0 X
4.6.4 Proof of Corollary 4.3

We start proving that when α1 is not too close of 0 or 1, the minimizers of F in
X are not radially symmetric. We show that for any radially symmetric ϕ ∈ X,
F(ϕ) > F(ϕds ), where the support of ϕds ∈ X is a disk sector. We first prove this
for functions such that {ϕ = 0} or {ϕ = π} is an annulus. Then, we generalize
by induction the result to functions ϕ such that {ϕ = 0} is composed of a finite
number of connected components. We conclude then by approximating any radially
symmetric ϕ ∈ X by this kind of functions.
We recall that ρ is given in (4.1.2) and that X is the space of functions ϕ ∈

BVloc (D; {0, π}) such that
Z
ρ = α1 . (4.6.17)
{ϕ=0}
Then, if ϕds ∈ X is such that {ϕds = 0} is a disk sector we easily compute

F(ϕds ) 3
= .
8σ 16
For 0 ≤ R− ≤ R+ ≤ λ we denote A(R− , R+ ) the annulus of center
R the origin, inner
− +
radius R and outer radius R . If ϕα = 1A(0,Rα ) is such that A(0,Rα ) ρ = α, then
√
Rα = λ(1 − 2)1/2 and
F(ϕα )
= f (α) ,
8σ
√
where f : [0, 1] → R+ is the concave function f (α) = α3/4 (1 − α)1/2 . We see that
there exists δ0 ∈ (0, 1/2) such that if α ∈ [δ0 , 1 − δ0 ], then f (α) > 3/16.
Proposition 4.8. If α1 ∈ [δ0 , 1−δ0 ], then the minimizers of F in X are not radially
symmetric functions.
Proof : (Step 1) Let ϕd1 ∈ X such that {ϕd1 = 0} = A(0, R). After (4.6.17), we have
that F(ϕd1 )/8σ = f (α1 ) so
F(ϕd1 ) > F(ϕds ) . (4.6.18)

Since α2 = 1 − α1 , α2 ∈ [δ0 , 1 − δ0 ] and the similar inequality holds if {ϕd1 = 0} =
A(R, λ). Let ϕa1 ∈ X such that {ϕa1 = 0} = A(R1 , R2 ), with 0 < R1 < R2 < λ.
Writing
Z Z
β1 = ρ and β2 = ρ.
A(0,R1 ) A(R1 ,R2 )
we compute
F(ϕa1 )
= f (β1 ) + f (β1 + β2 ) .
8σ
After (4.6.17), we have that β2 = α1 , so
F(ϕa1 )
= f (β1 ) + f (β1 + α1 ) .
8σ
The right hand size of the previous equality is a concave function of β1 and the value
F (ϕa )
of β1 may vary between 0 and α2 . If β1 = 0 then 8σ1 = f (α1 ), and if β1 = α2 then
F (ϕa
1)
8σ
= f (α2 ). We derive
F(ϕa1 ) > F(ϕds ) . (4.6.19)

(Step 2) Let ϕn ∈ X such that
n
[
{ϕn = 0} = A2j ,
j=1
− +
with A2j = A(R2j , R2j ) and
− + − +
0 ≤ R2j−2 < R2j−2 < R2j < R2j ≤λ
R R R
for 2 ≤ j ≤ n. We write β2j = A2j ρ, β1 = A(0,R− ) ρ, β2n+1 = A(R+ ,λ) ρ and
2 2n
Z
β2j+1 = ρ
+ −
A(R2j ,R2j+2 )
for 1 ≤ j ≤ n − 1. Notice that we allow A(0, R2− ) or A(R2n

+
, λ) to be empty, but this
only implies that β1 = 0 or β2n+1 = 0. With this notation we have
n
X n
X
β2i = α1 , β2i+1 = α2 (4.6.20)
i=1 i=1
and
j
2n
!
F(ϕn ) X X
= f βi =: gn (β1 , · · · , β2n ) .
8σ j=1 i=1
By induction, we are going to prove the following property :
n
X n
X
(Pn ) ∀ β1 , · · · , β2n+1 ∈ [0, 1] such that β2i = α1 et β2i+1 = α2 ,
i=1 i=1
F(ϕds )
gn (β1 , · · · , β2n ) > .
8σ
If n = 1 we are in one of the cases analyzed in Step 1, so (4.6.18) and (4.6.19) yield
(P1 ).
Let us assume that (Pn ) holds and consider β1 , · · · , β2n+3 ∈ [0, 1] such that
n+1
X n+1
X
β2i = α1 and β2i+1 = α2 . (4.6.21)
i=1 i=1
We have
j
2n
! 2n+1
! 2n+2
!
X X X X
gn+1 (β1 , · · · , β2n+2 ) = f βi +f βi +f βi .
j=1 i=1 i=1 i=1
The right hand side of previous equality Pn−1is a concave function of β2n+2 . The value of
β2n+2 may vary between 0 and α1 − i=1 β2i . Suppose first that β2n+2 = 0. Then,
defining β̃j = βj for j = 1, · · · , 2n and β̃2n+1 = β2n+1 +β2n+3 , the β̃i ’s satisfy (4.6.20)
and we have
j
2n
!
X X
gn+1 (β1 , · · · , β2n+2 ) ≥ f βi = gn (β̃1 , · · · , β̃2n ) .
j=1 i=1
Suppose now that β2n+2 = α1 − n−1

P
i=1 β2i . After (4.6.21) this implies β2n = 0. Then,
defining β̃j = βj for j = 1, · · · , 2n − 2, β̃2n−1 = β2n−1 + β2n+1 , β̃2n = β2n+2 and
β̃2n+1 = β2n+3 , the β̃i ’s satisfy (4.6.20) and we have
j
2n−2
! 2n−2
!
X X X
gn+1 (β1 , · · · , β2n+2 ) ≥ f βi +f βi + (β2n−1 + β2n+1 )
j=1 i=1 i=1
2n−2
!
X
+f βi + (β2n−1 + β2n+1 ) + β2n+2
i=1
= gn (β̃1 , · · · , β̃2n ) .
F (ϕds )
In both cases (Pn ) yields gn+1 (β1 , · · · , β2n+2 ) > 8σ
, so the result holds for all the
possible values of β2n+2 .
We have proved that if ϕn ∈ X is a radially symmetric function which support has

a finite number of connected components, then
F(ϕn ) > F(ϕds ) . (4.6.22)

(Step 3) Suppose now that ϕ ∈ X is a radially symmetric function such that {ϕ = 0}
has an infinite number of connected components. Since ϕ has locally finite perimeter
in D, {ϕ = 0} is the union of a countable family of disjoints annulus. We write
[
{ϕ = 0} = A2j
j∈Z
− +
with A2j = A(R2j , R2j ) such that
− + − +
0 < R2j < R2j < R2j+2 < R2j+2 < λ. (4.6.23)
For every n ∈ N, we define a function ϕn : D → {0, π} by
n
[ [ [
{ϕn = 0} = A2j Ã2n+2 Ã−2n−2 ,
j=−n
such that
Ã2n+2 = A(L+
n , λ) and Ã−2n−2 = A(0, L−
n)
with L− + − +
n , Ln to be chosen next. If (Ln , Ln ) = (0, λ), then (4.6.23) gives
Z n Z
X Z
ρ= ρ< ρ.
{ϕn =0} j=−n A2j {ϕ=0}
− +
Similarly if (L− +
n , Ln ) = (R−2n , R2n ), then
Z XZ XZ X Z Z
ρ= ρ+ ρ+ ρ> ρ.
+ − + −
{ϕn =0} j∈Z A2j j≥n A(R2j ,R2j+2 ) j≤−(n+1) A(R2j ,R2j+2 ) {ϕ=0}
− +
Hence,
R by continuity
R there is a pair (L− +
n , Ln ) ∈ (0, R−2n ) × (R2n , λ) such that
{ϕn =0}
ρ = {ϕ=0} ρ = α1 . Clearly ϕn ∈ BVloc (D), so ϕn ∈ X. Moreover, (4.6.23)
yields
lim L−
n = 0 and lim L+
n = λ. (4.6.24)
n→∞ n→∞
We have
XZ
ρ /2 dH1 ,
3
F(ϕ) =
+
j∈Z ∂B(0,R2j )
and since ρ is radially symmetric
n Z
X
− −
ρ /2 dH1 + 2π ρ /2 (L+ +
3 3 3/2

F(ϕn ) = n ) Ln + ρ (Ln ) Ln .
+
j=−n ∂B(0,R2j )
After (4.6.24), the last term in the previous equality goes to zero as n → +∞, so
limn→∞ F(ϕn ) = F(ϕ). Finally, since {ϕn = 0} has a finite number of connected
components, (4.6.22) yields F(ϕ) > F(ϕds ), which ends the proof.
Proof of Corollary 4.3 : this follows from Corollary 4.3 and Proposition 4.8.
4.7 Appendix
We end this article given the proof of Lemma 4.5, which is essentially the same
of Lemma 4.3 in [40], which in turn is a generalization of Lemma 1 in [130]. For
completeness we give here the details of the proof.
Proof of Lemma 4.5 : (Step 1) Suppose first that D ∩ A and D\A have both non
empty interior and let
B(x1 , δ) ⊂ D ∩ A and B(x2 , δ) ⊂ D\A . (4.7.1)

We first approximate A by sets of finite perimeter in D. For k ≥ 2 we define Dk =
D ∩ B(0, λ(1 − 1/k)) and A0k = A ∩ Dk . We have that ∂∗ A0k ⊂ (∂∗ A ∩ Dk ) ∪ ∂Dk , so
Z Z Z
3/2 1 3/2 1
ρ /2 dH1 .
3
ρ dH ≤ ρ dH +
∂∗ A0k ∂∗ A∩Dk ∂Dk
Using Lebesgue dominated convergence

R theorem, the first term in the right hand
side of the inequality converges to ∂∗ A ρ3/2 dH1 < +∞. The definition of Dk and
(4.2.8) yield
Z
3/2
2λ2 3/2
ρ /2 dH1 ≤ kρkL∞ (∂Dk ) H1 (∂Dk ) ≤ H1 (∂D) = ok→∞ (1) .
3
∂Dk k
Hence,
4.7. APPENDIX 129
Z Z
3/2 1
ρ /2 dH1 .
3
lim ρ dH ≤ (4.7.2)
k→∞ ∂∗ A0k ∂∗ A
(Step 2) Since A0k has finite perimeter in D, it can be approximated (see the proof
of Lemma 1 in [130]) by open bounded sets Ãk , such that
1
Λ2 (Ãk ∆A0k ) ≤ (4.7.3)
k
0
Ak ⊂ Ãk + B(0, /k)
1 and Ãk ⊂ A0k + B(0, 1/k) (4.7.4)
1
H (∂ Ãk ∩ ∂D) = 0 . (4.7.5)
The definition of A0k and (4.7.3) imply (i). Using (4.7.1) and (4.7.4), for large enough
k we get
B(x1 , δ) ⊂ Ãk and B(x2 , δ) ⊂ D\Ãk . (4.7.6)

Moreover, using (ii) from Proposition 2.3 in [40] and the fact that Ãk belongs to a
sequence Ãnk such that kÃnk kBV (D) → kA0k kBV (D) as n → 0, we have
Z Z
3/2 1
ρ /2 |D1A0k | + .
3
ρ |D1Ãk | ≤ (4.7.7)
D D k
Also, the definition of A0k and (4.7.3) yield
Z Z
γk := ρ− ρ = ok→∞ (1) . (4.7.8)
Ãk A
(Step 3) Now, we set


 Ãk \B(x1 , r1,k ) if γk > 0
Ak = Ãk if γk = 0 ,
Ãk ∪ B(x1 , r2,k )

if γk < 0
where r1,k and r2,k are chosen to satisfy
Z Z
ρ= ρ = γk .
B(x1 ,r1,k ) B(x2 ,r2,k )
R
Since r 7→ B(x1,2 ,r) ρ is continuous and decreasing for r ∈ (0, δ), r1,k and r2,k are
unique and tend to zero as k → ∞. Then, we derive from (4.7.6) and (4.7.8), for
large enough k, that
Z Z Z
ρ= ρ − γk = ρ.
D∩Ak Ãk A
Moreover, from (4.7.5) and (4.7.6), we have H1 (∂Ak ) = 0 for k large enough, so we
have proved (ii). Using again (4.7.6) we obtain
Z Z
3/2
ρ /2 |D1Ãk | + kρk∞ H 1 (∂B(x1 , r1,k ) ∪ ∂B(x2 , r1,2 )) .
3
ρ |D1Ak | ≤ (4.7.9)
D D
Hence, using (4.7.7), we obtain

Z Z
3/2
ρ /2 |D1A0k | + ok→∞ (1) ,
3
ρ |D1Ak | ≤
D D
so (4.7.2) gives
Z Z
3/2
ρ /2 |D1A |
3
lim sup ρ |D1Ak | ≤
k→∞ D D
and (iii) is proved.
(Step 4) We now remove the condition that D ∩ A and D\A have no empty interior.
First, we notice that Λ2 (D ∩ A) = 0 and Λ2 (D\A) = 0 are not possible because of
the mass constraints in (4.1.7). Hence, there exists x1 a point of density of D ∩ A
and x2 a point of density of D\A. Consider the function
Z Z
Φ(δ1 , δ2 ) = ρ− ρ,
A12 A
where A12 = A ∪ B(x1 , δ1 )\B(x2 , δ2 ). Since ρ > 0 in D, for any δ > 0 we have
Φ(δ, 0) > 0 and Φ(0, δ) < 0 .

Since Φ is continuous, there is t = tδ ∈ (0, 1) such that Φ(tδ, (1 − t)δ) = 0. Define
Aδ = A ∪ B(x1 , (1 −R t)δ)\B(xR 2 , tδ) and ϕ = π1 Aδ . Both D ∩ Aδ and D\Aδ have no
2
empty interior and Aδ ρ = A ρ. Moreover Λ (Aδ ∆A) → 0 as δ → 0, and using an
inequality similar to (4.7.9), we get
Z Z
3/2
ρ /2 |D1A | .
3
lim sup ρ |D1Aδ | ≤
δ→0 D D
Finally, for each Aδ we apply the construction from steps 1-3 and conclude thanks
to a diagonal argument, see Corollary 1.16 in [25].
Acknowledgements The second author would like to acknowledge discussions with

Guy Bouchitté, Pierre Seppecher and Duvan Henao. We would like to thank Clément
Gallo.
Chapitre 5
Spectrum of a magnetic Schrödinger

operator with a kagome potential
Le travail présenté dans ce chapitre a beaucoup

bénéficié de discussions avec P. Kerdelhue.
5.1 Introduction
In this chapter, we study in a semi-classical regime the selfadjoint extension Ph,A,V in
L2 (R2 ), of the Schrödinger operator with periodic magnetic potential A and periodic
electric potential V , defined in C0∞ (R2 ) by
0
Ph,A,V = (hDx1 − A1 (x))2 + (hDx2 − A2 (x))2 + V (x) . (5.1.1)
Here Dxj = 1i ∂xj and h > 0 is a “small” parameter.
In the quantum mechanics formalism, the operator Ph,A,V describes a charged par-
ticle in a two dimensional media under a transverse magnetic field. The spectrum
of the operator Ph,A,V corresponds to the possible values for the energy of the par-
ticle. The last years the search for artificial magnetism with ultra cold atoms (see
[65] and [102]) has renewed the interest in this physical system. We mention that
a two-dimensional kagome lattice for ultra cold atoms has been recently achieved
(see [106]). In this chapter, our goal is to study the spectrum of Ph,A,V as a function
of h and the magnetic field, when both the magnetic vector potential A and the
electric potential V are invariant by the symmetries of the kagome lattice. The main
motivation is to understand, and to analyze mathematically, various considerations
of Hou in [99].
In the case when A = 0 (see for example Chapter XIII.16 in [152]) the spectrum
of Ph,A,V is continuous and composed of bands. The general case, even when the
magnetic field is constant, is very delicate. Depending on the flux of the magnetic
field trough a unit cell, the spectrum can indeed become very singular (Cantor struc-
ture). To approach this problem, we are often lead to the study of limiting models in
different asymptotic regimes, such as discrete operators defined over `2 (Z2 ) or `2 (Z)
and pseudo-differential operators defined over L2 (R2 ).
131
132 CHAPITRE 5. THE MAGNETIC SCHRÖDINGER OPERATOR
To each basis B = {ν1 , · · · , νn } of Rn , we can associated a lattice Γ on Rn spanned

over Z by the vectors of B, that is,
Γ = α1 ν1 + · · · + αn νn ; (α1 , · · · , αn ) ∈ Zn .

The set of points Γ is called a Bravais lattice. Clearly, a Bravais lattice can be
generated by different basis. A fundamental domain of a Bravais lattice, which is
not unique, is a domain of the form

V = t1 ν1 + · · · + tn νn ; t1 , · · · , tn ∈ [0, 1] ,
where {ν1 , · · · , νn } is one of the basis generating Γ.
In dimension n = 2, if the magnetic vector potential A = (A1 , A2 ) is associated with

the 1-form
ωA = A1 dx1 + A2 dx2 ,
the magnetic field B is associated with 2-form obtained by taking the exterior deri-
vative of ωA :
dωA = B(x)dx1 ∧ dx2 .
The flux of B through a fundamental domain V of the lattice Γ is then given by
Z
η= dωA .
V
The spectrum of Ph,A,V depends on the normalized flux of the magnetic field through
a fundamental domain of the lattice associated with V , given by
η
γ= .
h
The spectrum of Ph,A,V have been studied for several two dimensional lattices. The
most simple one is the square lattice, which leads to the famous Hofstadter butterfly
when plotting the spectrum as a function of the normalized magnetic flux through
to a fundamental domain of the lattice (see Figure 5.2). We now briefly explain the
cases of the square, triangular and hexagonal lattices.
(a) (b)
(c) (d)
Figure 5.1 – (a) Square, (b) triangular, (c) hexagonal and (d) kagome lattices.
Square lattice. The square lattice is the Bravais lattice associated with the basis
{(1, 0), (0, 1)} of R2 (see Figure 5.1a). Each point of the lattice has 4 nearest neigh-
bors. One of the models used for the description of a charged particle in a square
lattice is the discrete operator Lγ defined on `2 (Z2 ) by
Lγ = τ1 + τ1∗ + t(τ2 + τ2∗ ) , (5.1.2)

where t > 0 is the tunneling parameter,
(τ1 v)n,m = vn−1,m , (τ2 v)n,m = e2iπγn vn,m−1 (5.1.3)

and γ is the normalized flux of the magnetic field through a fundamental domain of
the square lattice.
This operator was named “almost-Mathieu” by B. Simon, in reference to the Ma-

d2
thieu operator defined over the real line by − dx 2 + cos x .
Using a partial Floquet theory 1 (see [152], Chapter XII.16), we are lead to the study
of the spectrum of a family (parametrized by θ2 ) of discrete Schrödinger operators
Lγ,θ2 acting over `2 (Z) by
(Lγ,θ2 v)n = vn+1 + vn−1 + Vθ2 (n)vn , (5.1.4)

where Vθ2 (n) = 2t cos (2π(γn + θ2 )) is the potential.
1. Also called Bloch’s theory.

Indeed, introducing the Floquet condition τ2 v = ei2πθ2 v we have

[
σ(Lγ ) = σ(Lγ,θ2 ) ,
θ2 ∈[0,1]
where σ(L) denotes the spectrum of L and

n
D(Lγ,θ2 ) = v ∈ `2 (Z2 ) ; τ2 v = ei2πθ2 v} .
When γ is irrational, the spectrum of Lγ,θ2 does not depend on θ2 . This is no longer
the case when γ is rational. This is important to keep in mind when one tries to
understand the spectrum in the rational case as a limit of the spectra obtained in
the irrational case.
In 1976 Hofstadter performed a formal study of the spectrum of Lγ,θ2 as a func-

tion of γ ∈ Q ([98]). Its approach suggests a fractal structure for the spectrum and
leads to Hofstadter’s butterfly. The method consists in studying numerically the case
γ = p/q, with p, q ∈ N relative primes and p < q ≤ 50. Hofstadter found that in
this case, the spectrum is formed of q disjoint bands, which can only touch at their
boundary. Hofstadter’s butterfly is obtained by placing in the y-axis of a graph the
bands of the spectrum (see Figure 5.2). Moreover, Hofstadter derived rules for the
configuration of the bands related to the expansion of p/q as continued fraction.
This configuration suggests the Cantor stucture of the spectrum of Lγ,θ2 when γ is
irrational.
A longtime open problem, proposed by Kac and Simon in the 80’s and called the
“Ten Martinis problem” ([159], Problem 4), was to prove that for all irrational γ
and t > 0, the spectrum of Lγ is a Cantor set. After many efforts starting with the
article of Bellissard and Simon in 1982 ([30]), the problem was finally solved in 2009
by Avila and Jitomirskaya ([26]).
In order to compute numerically the spectrum of Lγ when γ = p/q, we may use

again the Floquet theory. Introducing the Floquet condition
vn+q = ei2πθ1 q vn ,
we are lead to the computation of the eigenvalues of the family (parametrized by θ1
and θ2 ) of hermitian matrices in Mq (C) defined by
M
p,q,θ1 ,θ2 = e
i2πθ1
K + e−i2πθ1 K ∗ + ei2πθ2 Jp,q + e−i2πθ2 Jp,q
∗
,
where
Jp,q = diag(exp (2iπ(j − 1)p/q))

and

1 if j = i + 1 (mod q)
Kij = .
0 if not
Denoting for γ = p/q the spectrum of Lγ by σγ we obtain
[
σγ = σ(M
p,q,θ1 ,θ2 ) ,
θ1 ,θ2 ∈[0,1]
and we observe the following properties (see Figure 5.2) :
σγ+1 = σγ (translation invariance) (5.1.5)

σ−γ = σγ (reflexion with respect to the axis γ = 0) (5.1.6)
−e ∈ σγ ⇔ e ∈ σγ (reflexion with respect to the axis e = 0) (5.1.7)
σγ ⊂ [−4, 4] (range of the spectrum) . (5.1.8)
We see that properties (5.1.5) and (5.1.6) follow from the definition of the operator
Lγ in (5.1.2) and the discrete translations τ1 , τ2 in (5.1.3).
We also notice that (5.1.5) together with (5.1.6) imply
σ1−γ = σγ (reflexion with respect to the axis γ = 1/2) . (5.1.9)

Coming back to the study of the operator Lγ , it is also possible to show that it is
unitary equivalent with the pseudo-differential operator defined in L2 (R) by
f (x + γ) + f (x − γ)
Lγ f (x) = + cos (x) f (x) . (5.1.10)
2
This operator corresponds to the γ-quantification of the symbol
p(x, ξ) = cos x + cos ξ .

We refer to the introductory notes [115] of Lein about Weyl quantification and to
the book [19] of Alinhac and Gérard for the theory of pseudo-differential operators.
Helffer and Sjöstrand developed in [94], [96] and [97] sophisticated techniques ins-
pired by the work of the physicist Wilkinson ([168]) to study the operator Lγ . In
particular, they justified in various regimes the approximation for the low spectrum
of Ph,A,V by the spectrum of Lγ .
Triangular lattice. The √

triangular lattice 2 is the Bravais lattice associated with
the basis {(1, 0), (1/2, − 3/2)} (see Figure 5.1b). Each point of the lattice has 3 nearest
neighbors. This case was studied by the physicists Claro and Wannier in [55]. These
authors exhibit an analogous structure to the case of the square lattice. In the case
γ = p/q, with p, q ∈ N relative primes, the spectrum is formed of q disjoint bands,
which can only touch at their boundary (see Figure 5.3). In [111], Kerdelhue studied
rigorously the operator Ph,A,V in this case, using the techniques developed by Helffer
and Sjöstrand. He justified the reduction to a pseudo-differential operator L4 γ defined
2
in L (R) corresponding to the γ-quantification of an analytic periodic symbol. The
principal symbol of the operator L4 γ is given by
2. We remark that the triangular and hexagonal lattices are sometimes respectively called hexa-
gonal and honeycomb lattices.
Figure 5.2 – Original Hofstadter’s butterfly (square lattice).
q(x, ξ) = cos x + cos ξ + cos(x − ξ) . (5.1.11)

Kerdelhue also showed that the γ-quantification of the symbol q(x, ξ) here before
has the same spectrum that the operator Qγ defined on L2 (Z2 ) by
X
Qγ f (α) = e−i2πγ(α1 +β1 )(α2 −β2 ) f (β) . (5.1.12)
β∈{(±1,0),(0,±1),(1,±1)}
We remark that the summation here before runs over the coordinates of the nearest
neighbors of a point in the triangular lattice.
In the case when γ = p/q the spectrum can be numerically computed by considering
the family of matrices in Mq (C) defined by
M4
p,q,θ1 ,θ2 = e
i2πθ1
K + e−i2πθ1 K ∗ + ei2πθ2 Jp,q + e−i2πθ2 Jp,q
∗
+ei2π(θ1 +θ2 ) KJp,q + e−i2π(θ1 +θ2 ) K ∗ Jp,q

∗
.
Denoting for γ = p/q the spectrum of Qγ by σγ , we obtain that σγ verifies (5.1.6)
and
e ∈ σγ+1 ⇔ −e ∈ σγ (translation anti-invariance) (5.1.13)

We notice that (5.1.13) together with (5.1.6) imply
e ∈ σ1−γ ⇔ −e ∈ σγ (anti-reflexion with respect to the axis γ = 1/2) . (5.1.15)
Property (5.1.13) follows from the definition of Qγ in (5.1.12). Indeed, replacing γ

by γ + 1, we obtain a new operator Q̃γ defined on L2 (Z2 ) by
X
Q̃γ f (α) = e−i2πγ(α1 +β1 )(α2 −β2 ) (−1)(α1 −β1 )(α2 −β2 ) f (β) ,
β∈{(±1,0),(0,±1),(1,±1)}
which corresponds to the γ-quantification of the symbol
q̃(x, ξ) = cos x + cos ξ − cos(x − ξ) .

Using the fact that the spectrum of the γ-quantification of a symbol is invariant
by the affine symplectic transformation (x, ξ) 7→ (x + π, ξ + π), we obtain that the
spectrum of the γ-quantifications of the symbols q(x, ξ) and −q(x, ξ) are the same.
This yields (5.1.13).
Figure 5.3 – Hofstadter’s butterfly for the triangular lattice.
Hexagonal lattice. The hexagonal lattice is not a Bravais lattice, but it corres-
ponds to a triangular lattice with two points per fundamental domain (see Figure
5.1c). Each point of the lattice has 3 nearest neighbors. This case was also rigorously
studied by Kerdelhue in [111] and [112]. We remark that this configuration corres-
ponds to a charged particle in a graphene sheet submitted to a transverse magnetic
field (see Chapter 6 in [134]). The work of Kerdelhue acquires a new interest after the
2010 Nobel Prize in Physics awarded to Geim and Novoselov for their experiments
involving graphene ([78], [139], [141]). In the case of a hexagonal lattice, Kerdel-
hue justified the reduction to a system of pseudo-differential operators defined in
[L2 (R)]2 which principal symbol is given by

0 1 + eix + eiξ
s(x, ξ) = . (5.1.16)
1 + e−ix + e−iξ 0
In the case when γ = p/q, the spectrum can be numerically computed by diagona-
lizing the hermitian matrices in M2q (C) defined by
 

 0q Iq + eiθ1 K + e−iθ2 Jp,q
∗ 

M7
 
p,q,θ1 ,θ2 =

.

 
 Iq + e−iθ1 K ∗ + eiθ2 Jp,q 0q 
The spectrum is formed of 2q disjoint bands, which can only touch at their boundary
(see Figure 5.4).
The spectrum verifies properties (5.1.5), (5.1.6) and (5.1.7) of the square lattice case
and (5.1.14) of the triangular lattice case.
Kagome lattice. The kagome lattice is a triangular lattice with three points per
fundamental domain (see Figure 5.1d). Each point of the lattice has 4 nearest neigh-
bors. The word kagome (see [129]) means a bamboo-basket (kago) woven pattern
(me). It seems that the kagome lattice was named by the Japanese physicist K. Hu-
simi in the 50’s and that the first article considering kagome lattices was published
in 1951 by the I. Syôzi in [162] (see Figure 5.5).
Considering the previous discussion about the square, triangular and hexagonal
lattices, in the kagome lattice case one may conjecture a reduction of the operator
Ph,A,V to a system of pseudo-differential operators defined in [L2 (R)]3 , corresponding
to the γ-quantification of a symbol of the form
 
0 r12 (γ, x, ξ) r13 (γ, x, ξ)
∗
r(γ, x, ξ) =  r12 (γ, x, ξ) 0 r23 (γ, x, ξ)  ,
∗ ∗
r13 (γ, x, ξ) r23 (γ, x, ξ) 0
∗
with rij (γ, ·, ·) being a trigonometric polynomial in x and ξ and where rij ,1≤i<
j ≤ 3, is the complex conjugate of rij .
In the case when the normalized flux trough a fundamental domain of the kagome
lattice is γ = p/q, one may expect to numerically compute the spectrum by diago-
nalizing a family of hermitian matrices in M3q (C) of the form
Figure 5.4 – Hofstadter’s butterfly for the hexagonal lattice.
 
 0q M12
p,q,θ1 ,θ2 M13
p,q,θ1 ,θ2 


 
 
 
∗
Mkag
 12 23

p,q,θ1 ,θ2 =
 Mp,q,θ1 ,θ2 0q Mp,q,θ1 ,θ2 
,
 
 
 
 13 ∗ 23 ∗ 
 Mp,q,θ1 ,θ2 Mp,q,θ1 ,θ2 0q 
where Mijp,q,θ1 ,θ2 , 1 ≤ i < j ≤ 3, is written in terms of the matrices Kand Jp,q . We
may also expect that the spectrum is formed of 3q disjoint bands, which can only
touch at their boundary.
Outline of the chapter
In this chapter we study the kagome lattice and we perform the reduction of Ph,A,V
to different discrete models. We present the Hofstadter’s butterfly corresponding to
the kagome lattice and we study some of its symmetries.
(a) (b)
Figure 5.5 – (a) Japanese basket showing the kagome pattern and (b) draw of the
kagome lattice from Syôzi’s article published in 1951.
This chapter is organized as follows :
– In Section 5.2 we give the definition of the kagome lattice and we study its group
of symmetries G(Γ). We also construct a family of potentials invariant by these
symmetries, which have the additional property that they can in principle be ex-
perimentally realized using lasers.
– Section 5.3 is devoted to the semi-classical analysis of the low lying spectrum of
Ph,A,V . First of all, the harmonic approximation shows the existence of an ex-
ponentially small (with respect to h) band in which one part of the spectrum
(including the bottom) is confined. This part of the spectrum will be called the
low lying spectrum. The rest of the spectrum is separated by a gap of size h/C.
We analyze the restriction of Ph,A,V to the spectral space attached to the low lying
spectrum. Expressing this operator in a suitable orthonormal basis, which keeps
the symmetries of the lattice, we obtain a discrete operator on [`2 (Z2 )]3 , written
in terms of τ1 and τ2 .
– In Section 5.4, we use a partial Floquet theory to reduce the study of the discrete
operator on [`2 (Z2 )]3 to another discrete operator on [`2 (Z)]3 . When the flux is
rational, we use again a Floquet theory to further reduce the discrete operator to
a family of square matrices written in terms of the matrices K and Jp,q . We also
recall the link (see [94], [111]) with a semi-classical pseudo-differential operator
acting on [L2 (R)]3 , with a new effective semi-classical parameter being the norma-
lized flux of the magnetic field. We finish by numerically computing (under some
additional assumptions) the spectrum of these matrices, we present the equivalent
of Hofstadter’s butterfly in this case and we discuss their symmetries.
5.2. THE KAGOME LATTICE 141
5.2 The kagome lattice

5.2.1 Definition and labeling
We now study the properties of the kagome lattice. We first give a definition of the
lattice and we study its group of symmetries.
Let ν a vector in S1 and r be the rotation of angle π/3 centered at the origin. For
` ∈ Z/6Z, we introduce (see Figure 5.6) the vectors
ν` = r` (ν) . (5.2.1)
For ` ∈ {1, 3, 5}, we denote by Γ` the triangular lattice spanned by 2ν1 and 2ν2 and
translated by ν` . We use an upper case letter M for a point in Γ` and a lower case
letter m for the coordinates of the point M in the basis
B = {2ν1 , 2ν2 } . (5.2.2)

For α ∈ Z2 we write then
Mα,` = 2α1 ν1 + 2α2 ν2 + ν` ∈ Γ` . (5.2.3)

Let k : Z2 → Z2 be the map associated with the matrix in the basis B of the rotation
r, that is :

0 1
k(α1 , α2 ) = (α1 , α2 ) = (−α2 , α1 + α2 ) . (5.2.4)
−1 1
We remark that k satisfies
k 6 = idZ2 (5.2.5)
and
k(α) ∧ k(β) = α ∧ β , (5.2.6)

where ∧ is the cross product

α1 β1
α ∧ β = det .
α2 β2
The coordinates of ν` in B are k ` (1, −1), so the coordinates of the point Mα,` ∈ Γ`
defined in (5.2.3) are
1
mα,` = (α1 , α2 ) + k ` (1, −1) . (5.2.7)
2
We define the kagome lattice as the union of the three triangular lattices Γ1 , Γ3 and
Γ5 :
Γ = Mα,` ; α ∈ Z2 , ` ∈ {1, 3, 5} ,

(5.2.8)
This provide us with a labeling of Γ (see Figure 5.7). Depending on the situation,
we will give the properties of the kagome lattice in terms of the points Mα,` or in
terms of their coordinates mα,` .
Figure 5.6 – Vectors ν1 , ν3 , ν5 , 2ν1 and 2ν2 .
Figure 5.7 – Labeling of the kagome lattice.

5.2.2 Study of the symmetries

We consider the following affine maps : the rotation r, the reflexion s with respect
to the axis spanned by ν1 + ν2 and the translations along the vectors 2νj :
τj (x) = x + 2νj , j ∈ Z/6Z . (5.2.9)

We denote G(Γ) the subgroup of the affine group of the plane generated by r, s, τ1
and τ2 . The composition rules in G(Γ) are :
τj+3 = τj−1 (5.2.10) s ◦ τ1 = τ2 ◦ s (5.2.14)

τj ◦ τj+2 = τj+1 (5.2.11) s ◦ τ3 = τ6 ◦ s (5.2.15)
r ◦ τj = τj+1 ◦ r (5.2.12) s ◦ τ4 = τ5 ◦ s . (5.2.16)
s◦r = r−1 ◦ s (5.2.13)
For α ∈ Z2 we define
τ α = τ1α1 τ2α2 . (5.2.17)

The relations (5.2.10)-(5.2.12) and (5.2.14) imply that
rτ α = τ k(α) r (5.2.18)
and
sτ α = τ z(α) s , (5.2.19)
where z : Z2 → Z2 is the map associated with the matrix in the basis B of the
reflexion s :

0 1
z(α) = (α1 , α2 ) = (α2 , α1 ) . (5.2.20)
1 0
We remark that z satisfies
z 2 = idZ2 (5.2.21)
and
z(α) ∧ z(β) = −α ∧ β . (5.2.22)

The study of Γ shows the following properties :
Proposition 5.1. The kagome lattice is invariant by the maps in G(Γ) and for
every M, N ∈ Γ there exists g ∈ G(Γ) such that g(M ) = N .
Proof. Consider Mα,` ∈ Γ and β ∈ Z2 . We have
τ β (Mα,` ) = 2α1 ν1 + 2α2 ν2 + 2β1 ν1 + 2β2 ν2 + ν`

= Mα+β,` .
Moreover, the definition of the vectors νj gives

r(Mα,` ) = 2α1 ν2 + 2α2 ν3 + ν`+1

= 2(−α2 )ν1 + 2(α1 + α2 )ν2 + 2ν`+1 + ν`−2 (5.2.23)
= Mk(α)+k`+1 (1,−1) , `−2 .
Similarly, using (5.2.13) we have
s(Mα,` ) = 2α1 ν2 + 2α2 ν1 + s(r` (ν))

= 2α2 ν1 + 2α1 ν2 + r−` (s(ν))
= 2α2 ν1 + 2α1 ν2 + r−` (ν3 )
= 2α2 ν1 + 2α1 ν2 + 2r3−` (ν) + ν−`
= Mz(α)+k3−` (1,−1) , −` .
This proves the first assertion of the proposition. The second assertion follows easily
considering for any pair Mα,` , Mβ,j ∈ Γ the element gα,β,`,j = τ β ◦ rj−` ◦ τ −α which
gives
gα,β,`,j (Mα,` ) = Mβ,j .
Remark 5.1. From the proof of Proposition 5.1, we see that for every Mα,` ∈ Γ and
β ∈ Z2
τ β (Mα,` ) ∈ Γ` , r(Mα,` ) ∈ Γ`−2 , and s(Mα,` ) ∈ Γ−` (5.2.24)

where −` is the inverse of ` for the addition in Z/6Z.
The group of symmetries of the kagome lattice acts on the functions of the plane by
(gu)(x) = u(g −1 (x)) ∀ g ∈ G(Γ) . (5.2.25)

This defines a group action of G(Γ) on C ∞ (R2 ) which can be extended as an unitary
action on L2 (R2 ). We have
Supp(gu) = g(Supp(u)) . (5.2.26)

For each element g ∈ G(Γ) there exist unique α ∈ Z2 , ` ∈ {1, . . . , 6} and δ ∈ {0, 1}
g = τ α r ` sδ . (5.2.27)
We define
n o
+
G(Γ) = g ∈ G(Γ) ; g writes as (5.2.27) with δ = 0 (5.2.28)
and
n o
G(Γ)− = g ∈ G(Γ) ; g writes as (5.2.27) with δ = 1 . (5.2.29)
5.2.3 Construction of kagome potentials

Definition and assumptions
We call V : R2 → R a “kagome” potential if it satisfies the following assumption.
Hypothesis 5.2.
gV = V ∀g ∈ G(Γ) . (5.2.30)
For the semi classical analysis performed in Section 5.3, we also need assumptions
on the regularity and the minima of V .
Hypothesis 5.3. The potential V is a real nonnegative C ∞ function, invariant by

Γ, whose minima are located on the kagome lattice :
V ≥ 0 and V (x) = 0 if and only if x ∈ Γ , (5.2.31)

HessV (x) > 0 ∀x ∈ Γ . (5.2.32)
It is rather easy to define a kagome potential. For example, taking a non positive
radially symmetric function V0 ∈ C0∞ (R2 ) with a unique non degenerate minimum
at the origin, the function
X
V = gV0
g∈G(Γ)
is a kagome potential.
But it is also interesting to give explicit examples of kagome potentials in the class
of trigonometric polynomials. This leaves indeed open the possibility to realize ex-
perimentally these potentials with lasers (see for example [66] and [157]).
In order to construct this type of potential, we will first obtain a potential Ṽ whose
maxima are located in the kagome lattice, and then we will consider V = −Ṽ +kṼ k∞ .
Remembering the definitions of the vectors νj from (5.2.1), we denote by ν ⊥ the

vector deduced from ν by a rotation of π/2 (see Figure 5.8) and for j ∈ {1, 3, 5} we
define
√
µj = 3 νj⊥ . (5.2.33)
Figure 5.8 – Vectors ν1 , ν3 , ν5 , µ1 , µ3 and µ5 .
Towards kagome trigonometric potentials
The first natural idea for constructing such a potential is to consider the sum
V̂ (x) = cos2 (x · πµ1 + φ1 ) + cos2 (x · πµ2 + φ2 ) + cos2 (x · πµ3 + φ3 ) , (5.2.34)
with (φ1 , φ3 , φ5 ) chosen such that
cos2 (x · πµj−2 + φj−2 ) + cos2 (x · πµj+2 + φj+2 )

is maximal on Γj and cos2 (x · πµj + φj ) is zero there. Unfortunately, this does not
work. The problem is that we get additional points where V̂ is more than 2, and those
points form a hexagonal lattice (see the Figure 5.9). In order to solve this problem,
we

x·πµj +φj
replace in the definition of V̂ cos(x·πµ1 +φ1 ) by cos (x · πµj + φj )+2 cos 3
.
By doing this, we enhance the value of the maxima and reduce the other values.
A family of trigonometric potentials
We define the potentials Vj : R2 → R as

2
x · πµj + φj
Vj (x) = cos (x · πµj + φj ) + 2 cos (5.2.35)
3
where
3π 3π 3π
φ1 = − , φ3 = and φ5 = . (5.2.36)
2 2 2
After (5.2.33), each function Vj is constant in the direction of νj , and since |µj |2 = 3,
it is invariant by the translation along the vector µi . Moreover, defining
Ṽ = V1 + V3 + V5
and
V = −Ṽ + kṼ k∞ , (5.2.37)

we obtain a kagome potential with non degenerate local minima on the kagome
lattice (see Figure 5.10) .
Proposition 5.4. The function V defined in (5.2.37) satisfies (5.2.30) and (5.2.32)
and has local minima at the point of the kagome lattice.
Proof.
Proof of (5.2.30). We have to prove that Ṽ is invariant by s, r and the translation
τj . Let i, j ∈ {1, 3, 5}, i 6= j. The definition of µi gives
√ 3
3 νi⊥ = ± δij ,
νj · µ i = νj ·
2
where δij is the Kronecker delta. We derive
(τj Vi )(x) = Vi (τj−1 (x))

2
x · πµi + φi − 2νj · πµi
= cos (x · πµi + φi − 2νj · πµi ) + 2 cos
3
2
x · πµi + φi
= cos (x · πµi + φi ∓ 3πδij ) + 2 cos ∓ πδij
3
2
x · πµi + φi
= cos (x · πµi + φi ) + 2 cos
3
= Vi (x) ,
so τj V = V . The definitions of the vectors µi and the phases φi give
(rVi )(x) = Vi (r−1 (x))

2
x · πr(µj ) + φi
= cos (x · πr(µi ) + φi ) + 2 cos
3
2
−x · πµi−2 + φi
= cos (−x · πµi−2 + φi ) + 2 cos
3
2
x · πµi−2 + φi−2 + (φi − φi−2 )
= cos (x · πµi−2 + φi−2 + (φi − φi−2 )) + 2 cos
3
2
x · πµi−2 + φi−2
= cos (x · πµi−2 + φi−2 ) + 2 cos
3
= Vi−2 (x) ,
so rV = V .
Similarly, we have 3 s(µi ) = µ−i which yields

3. We recall that i ∈ Z/6Z, so −1 = 5 and −3 = 3.
(sVi )(x) = Vi (s−1 (x))

2
x · πs(µj ) + φi
= cos (x · πs(µi ) + φi ) + 2 cos
3
2
x · πµ−i + φi
= cos (x · πµ−i + φi ) + 2 cos
3
2
x · πµ−i + φi−2 + (φi − φ−i )
= cos (x · πµ−i + φ−i + (φi − φ−i )) + 2 cos
3
2
x · πµ−i + φ−i
= cos (x · πµ−i + φ−i ) + 2 cos
3
= V−i (x) ,
so sV = V . We conclude that gV = V for every g ∈ G(Γ).
We now prove that Ṽ have local maxima at the points of Γ. For each j ∈ {1, 3, 5},
the maxima of Vj are given by the condition (x · πµj + φj )/3 ≡ 0 (mod 2π). Hence,
the maxima are located in the lines

2 3
Lj,k = x ∈ R ; x · µj = (2k + 1) , k ∈ Z.
2
We observe numerically (see Figure 5.10) that maxima of Ṽ are located at the
intersections Lj,k ∩ Li,k0 for i 6= j, and k, k 0 ∈ Z. We claim that every point in the
sublattice Γj is located at the intersection of a line of maxima of Vj−1 with a line of
maxima of Vj+1 . Indeed, remembering the definition of Mα,j in (5.2.3) we compute :
Mα,j · µi = (2α1 ν1 + 2α2 ν2 + νj ) · µi

√ 3
= 2 3 (α1 ν1 + α2 ν2 ) · νi⊥ ± δij
 √
2
3
√  √
 2
α2 if i=1
3 3
= ± δij + 2 3 · 2
(−α
√ 1
− α2 ) if i=3
2  3

2
α1 if i=3
3
= (2kα,i ± δij )
2
with kα,i ∈ Z.
0 00
Hence, for every α ∈ Z2 , Mα,j ∈ Lj−1,k0 ∩ Lj+1,k00 for some integers k and k .
Moreover, the same computation gives that Vj (Mα,j ) = 0 and that
V (Mα,j ) = 18
for every α ∈ Z2 and j ∈ {1, 3, 5}.
Proof of (5.2.32). A straightforward computation gives

∇Vi (Mα,j ) = 0
and
π2
Hess Vi (Mα,j ) = (−22 δij − 8(1 − δij )) µi ⊗ µi .
3
Hence, HessṼ (M ) < 0, HessV (M ) > 0 for every M ∈ Γ.
Remarks
We conclude this section with three remarks.
Remark 5.5. Our numerical computation (see Figure 5.10) shows that the condition
(5.2.31) is verified but we do not have a mathematical proof.
Remark 5.6. We remark the potential defined by (5.2.37) with
p
x · πµj + φj
Vj (x) = cos (x · πµj + φj ) + 2 cos
3
where p ∈ 2N, is also a kagome potential (see Figure 5.11).
When p goes to +∞, we observe that the minima are very well localized at the points
of Γ. This could be an advantage for verifying theoretical assumptions for an accurate
semi-classical analysis of the tunneling effect between wells in the next section, but
p large is not experimentally reasonable.
Remark 5.7. Considering any Bravais lattice with three points by periodicity cell,
we are lead to the same situation, but the kagome lattice have a much richer struc-
ture.
(a) V1 (b) V3
(c) V5 (d) V
Figure 5.9 – The potentials Vj (x) = cos2 (x·µj +φj ) for j ∈ {1, 3, 5}, with µj defined
in (5.2.33) and φj in (5.2.36). The maxima of V = kV1 + V3 + V5 k∞ − (V1 + V3 + V5 )
are located in a hexagonal lattice. The minima of each potential are in dark blue
and the minima in dark red.
(a) V1 (b) V2
(c) V3 (d) V
Figure 5.10 – The potentials Vj from (5.2.35) for j ∈ {1, 3, 5}. The function V =
kV1 + V3 + V5 k∞ − (V1 + V3 + V5 ) is a kagome potential. The maxima of each potential
are in dark red and the minima in dark blue.
(a) p=2 (b) p=10
(c) p=30 (d) p=50

p/2 p/2 p/2
Figure 5.11 – The function V = V1 + V3 + V5 with p ∈ {2, 10, 30, 50} and Vj
p/2 p/2 p/2 p/2
given by (5.2.35) for j ∈ {1, 3, 5}. The function V = kV1 + V3 + V5 k∞ − (V1 +
p/2 p/2
V3 + V5 ) is a kagome potential. The maxima of each potential are in dark red
and the minima in dark blue.
5.3. THE SCHRÖDINGER MAGNETIC OPERATOR ON L2 (R2 ) 153
5.3 The Schrödinger magnetic operator on L2(R2)

5.3.1 The Schrödinger magnetic operator
In this section we adapt to the case of the kagome lattice a part of the work of
Helffer and Sjöstrand in [96] and Kerdelhue in [111]. The first authors were mainly
considering the case of a square lattice and Kerdelhue was analyzing the triangular
and hexagonal lattices.
0
We recall that we start from Ph,A,V defined in (5.1.1). Since we have assumed V ≥ 0,
∞
the operator is semi-bounded on C0 (R2 ) and there is an unique selfadjoint extension
in L2 (R2 ), which can be obtained as the Friedrichs extension of Ph,A,V 0
(see for
example [88]). It can be proved that the domain of Ph,A,V is given by
n o
D(Ph,A,V ) = u ∈ L2 (R2 ) ; Ph,A,V u ∈ L2 (R2 ) . (5.3.1)
The electric potential is supposed to satisfy Hypotheses 5.2 and 5.3.
The magnetic vector potential A = (A1 , A2 ) corresponds to the 1-form
ωA = A1 dx1 + A2 dx2 ,
that we sometimes simply write A. The magnetic field σB is the 2-form obtained by
taking the exterior derivative of ωA :
σB = dωA = B(x)dx1 ∧ dx2 .

In the case of R2 , we identify the 2-form σB with B.
The magnetic potential is supposed to satisfy the following assumptions :
Hypothesis 5.8. The potential A = (A1 , A2 ) is a C ∞ vector field such that the
corresponding magnetic 2 form satisfies
g(B dx1 ∧ dx2 ) = −B dx1 ∧ dx2 if g ∈ G(Γ)− (5.3.2)

and
g(B dx1 ∧ dx2 ) = B dx1 ∧ dx2 if g ∈ G(Γ)+ . (5.3.3)
We remark that these two conditions write gB = B for every g ∈ G(Γ). In particu-
lar, they are verified by a constant magnetic field.
5.3.2 Quantization of G(Γ)

The use of the symmetries in the case of the square, triangular and hexagonal lat-
tices was crucial in [94] and [111]. In order to take advantage of the properties of
the kagome lattice, we need to quantify the elements of G(Γ), that is, to associate
which each element of G(Γ) an unitary transformation in L2 (R2 ), which respects the
domain and commutes with Ph,A,V . These operators will be used later to study the
low lying spectrum of Ph,A,V . We note that the quantification of the translations τj
was introduced by Zak in [170]. We also mention the work of Helffer and Sjöstrand
([95], pages 147-148) who studied the case of constant magnetic field in arbitrarily
dimension and refer to Bellissard ([28]), Cartier ([52]) and Zak.
Quantization of the reflexion
The reflexion s ∈ G(Γ)− plays a particular role because of the change of sign appea-
ring in (5.3.2), which will lead us to construct an antilinear operator.
First, we try to quantify the reflexion s by S, defined over L2 (R2 ) by
Su (x) = u(s−1 (x)) . (5.3.4)

The operator S is clearly anti-unitary but does not necessarily commute with Ph,A,V ,
we have indeed
SPh,A,V = SPh,−sA,V .
This works only if A + sA = 0. We will perform a gauge transformation to get this
property. This is the object of the next lemma.
Lemma 5.9. There exists a real function φs ∈ C ∞ (R2 ) such that S commutes with
Ph,A− 1 dφs ,V .
2
i
The gauge transformation is defined on L2 (R2 ) by u 7→ e 2h φs u. We observe that
i i
e− 2h φs Ph,A,V e 2h φs = Ph,A− 1 dφs ,V
2
and that the magnetic field B is unchanged.
Proof. After (5.3.2), we have d(A + sA) = 0, so there is a real smooth function φs
on R2 such that
A + sA = dφs . (5.3.5)
Since s2 = idR2 , we obtain sA + A = d(sφs ), so

1 1
s A − dφs + A − dφs = 0 . (5.3.6)
2 2
Hence, defining
1
Ã = A − dφs
2
we get
Ã + sÃ = 0 . (5.3.7)
We have then
(−ihd − Ã)Su = −ihd(su) − Ã (su)

= ihs(du) − Ã (su)

= s ihdu − (S Ã) u .
So using (5.3.7) we get

(−ihd − Ã)Su = s ihdu + Ã u
= −S(−ihd − Ã)u ,
which gives the lemma.
From now on, we choose the gauge for A such that
A + sA = 0 (5.3.8)
so [Ph,A,V , S] = 0 is satisfied.
Remark 5.10. In the case of a constant magnetic field, the vector potential A(x1 , x2 ) =
B
2
(−x2 , x1 ) is the right gauge for the conclusion of Lemma 5.9.
Quantization of the rotation and the translations
We now quantify some of the elements in G(Γ)+ . For any g ∈ G(Γ)+ the 1-form
A − gA is closed and in fact it is exact. Indeed, by assumption (5.3.3),
d(A − gA) = dA − g dA = B − gB = 0 . (5.3.9)

Hence, there is a real smooth function φg , defined up to a constant, such that
A − gA = dφg . (5.3.10)
Later, we will use this freedom of choice of the constants to obtain simple commu-
tation properties.
We quantize then g ∈ G(Γ)+ by the operator Tg , defined on C0∞ (R2 ) by

i
(Tg u)(x) = e h φg (x) u(g −1 (x)) , (5.3.11)
where φg is the a real function associated with g by (5.3.10).
Lemma 5.11. Let g ∈ G(Γ)+ . The operator Tg is unitary on L2 (R2 ) and commutes
with Ph,A,V .
Proof. For the first assertion a simple computation gives

i −1 φ
Tg−1 = e− h g g
g −1 = Tg∗ . (5.3.12)
We have the equality between 1-forms
(−ihd − A)Tg u = ei
φg/h
((dφg ) gu − ih g(du) − A (gu)) ,
so using (5.3.10) we get
(−ihd − A)Tg u = ei g/h (−ih g(du) − (gA) (gu))

φ
= ei g/h g(−ihdu − Au)

φ
= Tg (−ihd − A)u ,
which gives the Lemma.
Definition of the magnetic rotation and translations
For j ∈ {1, 2, 3} we define the magnetic translations

i
Tj = e h φj τj , (5.3.13)
where φj are the real functions associated with τj by (5.3.10) with g = τj .
By (5.3.11) this gives

i
(Tj u)(x) = e h φj (x) u(τj−1 (x)) . (5.3.14)
The inverse of Tj is given by (5.3.12) and is also a magnetic translation.For j ∈

{4, 5, 6} we define
−1
Tj = Tj+3 . (5.3.15)
We also define a magnetic rotation

i
F = e h f r−1 , (5.3.16)
where f is the real function associated with r−1 by (5.3.10). By (5.3.11) this gives
i
(F u)(x) = e h f (x) u(r(x)) . (5.3.17)
Remark 5.12. We need a clever and explicit choice in order to be able to compare
Tg1 g2 and Tg1 ◦ Tg2 for any g1 , g2 ∈ G(Γ)+ . Hence we will only use the previous
construction for r, τj ∈ G(Γ)+ , j ∈ Z/6Z.
Remark 5.13. In the case of a constant magnetic field, choosing the gauge A(x1 , x2 ) =
B
2
(−x2 , x1 ) we have
f (x) = f0
B
φj (x) = − x ∧ (2νj ) + cj ,
2
where f0 and cj , j ∈ {1, 2, 3}, are arbitrarily constants.
Commutation rules
We now show how a good choice of the constants appearing in the definition of f
and φj lead to nice commutation rules for the operators F , S and Tj . Here we mainly
follow Section 1 of [111] and we complete the proof of Lemma 1.6 there.
Proposition 5.14. (i) The flux of B through a fundamental domain V of the lattice
generated by 2νj and 2νj+1 is independent of j ∈ Z/6Z and we write
Z
η= dωA .
V
(ii) We have
iη
Tj Tj+1 = e h Tj+1 Tj . (5.3.18)
(iii) There are unique φ1 , φ2 , φ3 and f such that
F 6 = IdL2 (R2 ) (5.3.19)

Tj F = F Tj+1 , (5.3.20)
and for these φ1 , φ2 and φ3 we have

iη
Tj Tj+2 = e 2h Tj+1 . (5.3.21)
(iv) Moreover,
SF = F ∗ S , (5.3.22)
and
ST1 = T2 S . (5.3.23)
We recall that for a path γ and a 1-form ω = ω1 dx1 + ω2 dx2 we define
Z Z h i
ω= ω1 (γ(t))γ10 (t) + ω2 (γ(t))γ20 (t) dt .
γ
Proof. Since the translations τj commute between them we have

i
Tj Tj+1 = e h {φj +τj φj+1 −φj+1 −τj+1 φj } Tj+1 Tj . (5.3.24)
After (5.3.9), the expression between the brackets in (5.3.24) is a constant. We
denote it by
i
ηj = {φj + τj φj+1 − φj+1 − τj+1 φj } . (5.3.25)
h
Z Z Z Z
(φj −τj+1 φj )(x) = dφj = (A−τj A) = A+ A,
[x,x−2νj+1 ] [x,x−2νj+1 ] [x,x−2νj+1 ] [x−2νj −2νj ,x−2νj+1 ]
(5.3.26)
where [x, y] denotes the path [0, 1] 3 t 7→ γ(t) = (1 − t)x + ty.
Similarly,
Z Z
(τj φj+1 − φj+1 )(x) = A+ A.
[x−2νj ,x] [x−2νj −2νj+1 ,x−2νj+1 ]
Hence, Stokes theorem yields

Z
ηj = dωA , (5.3.27)
Vj,j+1
where Vj,j+1 is a cell of periodicity of the lattice generated by 2νj and 2νj+1 with
vertex x, x − 2νj+1 , x − 2νj and x − 2νj − 2νj+1 .
After (5.3.3) the magnetic field Bdx1 ∧ dx2 is invariant by r, so the value of the ηj
do not depend on j, ` ∈ Z/6Z. This proves (i). Writing η = η1 we have then
i
Tj Tj+1 = e h η Tj+1 Tj , (5.3.28)
so (ii) is proved.
Since r6 = idZ2 we have

i 5f }
F 6 = e h {f +rf +···+r , (5.3.29)
and after (5.3.9) the expression between the brackets in (5.3.29) is a constant. Hence,
choosing an appropriate constant in the definition of f we obtain (5.3.19). For the
proof of relation (5.3.20), (5.2.10) and (5.2.12) give
i −1
T1 F = e h {φ1 +τ1 f −f −r φ2 } F T2
i −1
T2 F = e h {φ2 +τ2 f −f −r φ3 } F T3
i −1
T3 F = e h {φ3 +τ3 f −f +τ3 r φ1 } F T4 .
As before, after (5.3.9) the expressions between the brackets in the previous equali-
ties are constants. If we add the constants a1 , a2 and a3 , respectively to φ1 , φ2 and
φ3 , the expressions between the brackets are respectively modified by a1 − a2 , a2 − a3
and a1 + a3 . Hence, since the matrix
 
1 −1 0
 0 1 −1  (5.3.30)
1 0 1
is invertible, there exist a1 , a2 and a3 such that (5.3.20) is satisfied for j = 1, 2, 3.
Since Tj+3 = Tj−1 , (5.3.20) also holds for j = 4, 5, 6.
For the proof of relation (5.3.21), reasoning as before, {φj + τj φj+2 − φj+1 } is a
constant that we call c, and we have
c
Tj Tj+2 = ei h Tj+1 . (5.3.31)
i hc
Conjugating the previous equality by F and using (5.3.20), we obtain e Tj+2 =
Tj+1 Tj−1 , which gives
c
Tj Tj+2 = e2i h Tj+2 Tj . (5.3.32)
The proof of (5.3.18) also applies when taking Tj and Tj+2 instead of Tj and Tj+1 ,
i
so we have Tj Tj+2 = e h η Tj+2 Tj . Thus, 2c/h ≡ η/h [2π], which gives (c − η/2)/h ∈ πZ.
Since c and η do not depend on h, we derive that necessarily c = η/2, so (5.3.21) is
proved 4 . We have proved (iii).
Similarly, (5.2.13) yields
i
SF = e h {rf −sf } F ∗ S .
After (5.3.9) and (5.2.13) we have
d{rf − sf } = −(sA + A) + r(A + sA)
so (5.3.8) gives that rf − sf is a constant. In particular (rf − sf )(0) = 0, so we

obtain (5.3.22). Again, we have that
i
ST1 = e− h {sφ1 +φ2 } T2 S .
In view of (5.3.8), (5.3.9) and (5.2.14) we have
d{sφ1 + φ2 } = s(A − τ1 A) + (A − τ2 A)
= (A + sA) − τ2 (A + sA) = 0 ,
i
so ST1 = e h ĉ T2 S for some constant ĉ. Conjugating this equality by F 3 , (5.3.22)
i i
gives ST1−1 = e h ĉ T2−1 S, so ST1 = e− h ĉ T2 S, which yields ĉ ≡ 0 [2π]. We have proved
(iv).
4. This completes the proof of Lemma 1.6 in [111].

From now on, we choose φ1 , φ2 , φ3 and f in the definition of T1 , T2 , T3 and

F such that (5.3.19)-(5.3.23) are satisfied.
Remark 5.15. In the case of a constant magnetic field, choosing A(x1 , x2 ) =

B
2
(−x2 , x1 ) we verify that
f 0 = c1 = c2 = c3 = 0 .
Magnetic translation (continued)
For α ∈ Z2 we define the magnetic translation

iη
T α = e− 2h α1 α2 T1α1 T2α2 . (5.3.33)
We obtain the following relations :
Proposition 5.16. For every α, β ∈ Z2 we have
(T α )−1 = T −α (5.3.34)
iη
α β α∧β α+β
T T = e 2h T (5.3.35)
α k−1 (α)
FT = T F (5.3.36)
α z(α)
T S = ST . (5.3.37)
Proof. Using (5.3.18) we have
iη
(T α )−1 = e 2h α1 α2 T2−α2 T1−α1
iη
= e 2h (α1 α2 −2α1 α2 ) T1−α1 T2−α2
= T −α ,
and
iη
T α T β = e− 2h (α1 α2 +β1 β2 ) T1α1 T2α2 T1β1 T2β2
iη
= e− 2h (α1 α2 +β1 β2 +2α2 β1 ) T1α1 +β1 T2α2 +β2
iη
= e 2h (α∧β) T α+β .
iη
F T α = e− 2h α1 α2 F T1α1 T2α2
iη
= e− 2h α1 α2 T3−α1 T1α2 F
iη 2
= e− 2h (α1 α2 −α1 ) T1α1 T2−α1 T1α2 F
iη 2
= e− 2h (−α1 α2 −α1 ) T1α1 +α2 T2−α1 F
−1 (α)
= Tk F.
Finally, (5.3.37) is a direct consequence of (5.3.23).
5.3.3 Construction of a basis of the space attached to the low

lying spectrum of Ph,A,V
Preliminaries
In this subsection, in order to understand the analogous of the Hofstadter butterfly

appearing in connection with the kagome lattice, we study the low lying spectrum
of Ph,A,V . For this, we follow the method of “filling the wells” to obtain a basis of
the spectral space attached to the low lying spectrum. The wells are defined as the
connected parts of {x ∈ R2 ; V (x) ≤ 0}. In our setting, Hypothesis 5.3 yields that
the wells correspond to the points of the kagome lattice. The method consists in
associating with each well the Schrödinger operator obtained by filling all the other
wells. In this way, we obtain a family of one well Schrödinger operators. Then, the
idea is that, considering the space spanned by the ground states of all the one well
Schrödinger operators, we get a basis of the space attached to the low lying spec-
trum of Ph,A,V .
Therefore, we first study the one well case using the harmonic approximation. Then,
we present the Agmon distance related to the potential V , which is the main tool to
estimate the tunneling effect between the wells of Ph,A,V . Next, using the magnetic
rotation and translations, we construct a family of functions attached to the wells
of V . We estimates the tunneling effect and we give the action of the magnetic
operators over these functions. Finally, we project this family onto the space attached
to the low lying spectrum of Ph,A,V and we obtained the desired basis after an
orthonormalization process, which respects the action of the the magnetic operators.
The harmonic approximation
Here we recall a result from [93] about the semi classical analysis of the bottom of
the spectrum of a Schrödinger operator with magnetic field in the case when the
electric potential V has a unique non degenerate well at a point M .
The theory of the harmonic approximation was initially introduced for a Schrödinger
operator without magnetic field in [92] and [160] and can be extended to the magnetic
case (see Section 4.3.1 [86]). More precisely, the harmonic approximation consists in
replacing the potential V by its quadratic approximation at M and the magnetic
field by its value at M , that is the magnetic potential by its linear part at M . This
reads :
1
Phar,h,B = h2 Dx21 + (hDx2 − Bx1 )2 + hHessV (M ) x, xi . (5.3.38)
2
where B = B(M ) is the value of the magnetic field at the point M . The result is
the following.
Proposition 5.2. Assume that HessV (M ) > 0.The spectrum of the operator Phar,h,B
defined in (5.3.38) is discrete. The first eigenvalue is simple and given by
q
λhar,h,B = h λ21,0 + B 2 ,
√
where λ21,0 = (λ1 + λ2 )/ 2 is the first eigenvalue of Phar,1,0 and λ1 ,λ2 are the eigen-
values of HessV (M ).
A result of [93] allows then to estimates the first eigenvalue of a single well Schrö-
dinger operator using the harmonic approximation. We also refer to Chapter 11 of
[63] for other results in this spirit.
Proposition 5.3. Consider a vector field A = (A1 , A2 ) ∈ C ∞ (R2 ) and a real nonne-
gative potential V ∈ C ∞ (R2 ) having an unique non degenerate minimum at a point
M ∈ R2 . The smallest eigenvalue λh,B of the magnetic Schrödinger operator
Ph,A,Ṽ = (hDx1 − A1 (x))2 + (hDx2 − A2 (x))2 + Ṽ (x) (5.3.39)
is simple and satisfies
|λh,B − hλhar,1,B | ≤ Ch /2 .
3
There exists ε0 > 0 such that σ(Ph,A,Ṽ ) ∩ [0, h(λhar,1,B + ε0 )] = {λh,B }.
Remark 5.17. In the case of a weak constant magnetic field B = hB0 , the harmonic
approximation has no magnetic contribution and we have
|λh,B − hλhar,1,0 | ≤ Ch /2 .
3
Agmon estimates
Consider the Agmon metric V dx2 . For a piecewise C 1 curve γ, we can define its
length |γ| in this metric, and for x, y ∈ R2 we define the Agmon distance dV (x, y) as
the inf |γ| over all piecewise C 1 curves γ joining x to y. This distance may be dege-
nerate in the sense that dV (x, y) = 0 for x 6= y, but it satisfies standard properties
such as
dV (x, y) = dV (y, x) and dV (x, y) ≤ dV (x, y) + dV (y, z) .
For ϕ ∈ L2 (R2 ) and y ∈ R2 we use the notation
−dV (·,y)(1−ε)+ε

ϕ = Oε e h ,
which means that for every ε > 0, there exists hε > 0 and Cε such that
−d (·,y)(1−ε) ε
V h
e ϕ(·) ≤ Cε e h

L2 (R2 )
for h ∈ (0, hε ). Here dV (·, M ) is the Agmon distance to the point M . We refer to
Chapter 6 of the book of Dimassi and Sjöstrand [70] for more details on Agmon
estimates.
Carlsson’s construction of an orthonormal basis of the low lying spectrum
We now explain Carlsson’s construction of a orthonormal basis of the spectral space

attache to the low lying spectrum of Ph,A,V . The approach of Carlsson is quite gene-
ral (no assumption of periodicity is needed) but he does not consider the case with
magnetic field. Nevertheless, the theory is simpler in the periodic case and it was
shown in Section 9 of [94] how to generalize his result with the help of [93].
(Step 1 ) We start by considering δ ∈ (0, 1/8) and a nonnegative radial C0∞ function
χ such that χ = 1 in B(0, δ/2) and supp χ ⊂ B(0, δ). With every M ∈ Γ we associate
the operator
Pm = P + Vm ,
where
X
Vm (x) = χ(x − N ) .
N ∈Γ\{M }
We have seen in Proposition 5.1 that for all M, N ∈ Γ there is g ∈ G(Γ)+ such that
g(M ) = N . Considering the associated Tg defined in (5.3.12), all the operators Pm ,
M ∈ Γ, are unitary equivalent.
A result of Persson ([147]) gives that σ(Pm ) is discrete in the interval [0, b] where
b = lim inf (V + Vm )(x) .

|x|→∞
The operator Pm is a Schrödinger operator with an electric potential V + Vm . Using

Hypothesis 5.3, V + Vm has a unique non degenerate minimum, so we can apply
Proposition 5.3 to Pm . The first eigenvalue λh,B of Pm is simple and there exists
ε0 > 0 such that σ(Pm ) ∩ I(h) = {λh,B }, where I(h) = [0, h(λhar,1,B(M ) + ε0 )] with
B(M ) the value of the magnetic field at the point M .
(Step 2 ) Consider M1 = M(0,0),1 and let ϕ1 be an eigenfunction of Pm1 with eigen-

value λ(h) such that
kϕ1 kL2 (R2 ) = 1 and ϕ1 (M1 ) is real. (5.3.40)

For ` ∈ {1, 3, 5} we define
ϕ` = F 1−` ϕ1 , (5.3.41)
and for every Mα,` ∈ Γ we define an eigenfunction of Pmα,` with eigenvalue λ(h), by
iη 1 `
ϕmα,` = e− 2h α∧ 2 k (1,−1)
T α ϕ` . (5.3.42)
Defining rm = (P − λ(h))ϕm , we have the Agmon estimates
−dV (x,M )(1−ε)+ε
ϕm , rm , ∇A ϕm , ∇A rm = Oε e h , (5.3.43)
where ∇A = (hDxj − Aj (x))j=1,2 . Moreover,

[
supp rm ⊂ B(N, δ) . (5.3.44)
N ∈Γ\{M }
We also observe by the harmonic approximation that
ϕ1 (M1 ) = h− /2 char + O(1) .

1
(5.3.45)
We now give the action of the magnetic operators over the eigenfunctions ϕm .
Proposition 5.18. For every h > 0 there exist c ∈ {−1, 1} and d ∈ S 1 , such that
for all M ∈ Γ and β ∈ Z2 we have
iη
T β ϕm = e 2h β∧m ϕm+β (5.3.46)
F ϕm = c ϕk−1 (m) (5.3.47)
Sϕm = d ϕz(m) . (5.3.48)
Proof. Let M = Mα,` . For the first relation, using (5.3.35) we have
iη 1 `
T β ϕm = e− 2h α∧ 2 k (1,−1) T β T α ϕ`
iη 1 `
= e− 2h (α∧ 2 k (1,−1)−β∧α) T α+β ϕ `
iη iη
β∧m − 2h (α+β)∧ 12 k` (1,−1) α+β
= e 2h e T ϕ`
iη
β∧m
= e 2h ϕm+β .
For the second relation we first remark that
1
k −1 (mα,` ) = k −1 (α) + k `−1 (1, −1)
2
1
= k −1 (α + k ` (1, −1)) + k `−1 (−1, 1)
2
1
= k −1 (α + k ` (1, −1)) + k `+2 (1, −1) ,
2
so we have to prove that
iη −1 (α+k ` (1,−1))∧ 1 k `+2 (1,−1) −1 (α+k ` (1,−1))
F ϕm = c e− 2h k 2 Tk ϕ`+2 (5.3.49)
for some c ∈ {−1, 1}.
τ1 r−3 Vm1 = Vm1 ,

so after Lemma 5.11, T1 F 3 commutes with Pm1 . Hence, since λ(h) is a simple eigen-
value, there is a complex number c, such that |c| = 1 and
T1 F 3 ϕ1 = c ϕ1 . (5.3.50)
Moreover, (5.3.19) and (5.3.20) yield (T1 F 3 )2 = IdL2 (R2 ) , so c2 = 1.

−4 (−1,1)
F ϕ1 = c F 4 T1−1 ϕ1 = c T k F −2 ϕ1 = c T (1,−1) F −2 ϕ1 . (5.3.51)
This, together with (5.3.35), (5.3.36) and (5.2.5), gives
iη 1 ` −1 (α)
F ϕm = e− 2h α∧ 2 k (1,−1)
Tk F 1−` F ϕ1
iη
− 2h α∧ 12 k` (1,−1) −1 (α)
= ce Tk F 1−` T (1,−1) F −2 ϕ1
iη 1 ` −1 (α) k`−1 (1,−1)
= c e− 2h α∧ 2 k (1,−1)
Tk T1 F 1−(`+2) ϕ1
iη 1 ` iη −1 (α)∧k `−1 (1,−1) −1 (α+k ` (1,−1))
= c e− 2h α∧ 2 k (1,−1)
e 2h k Tk ϕ`+2 .
We remark that −k `−4 (−1, 1) = k `+2 (1, −1) and that using (5.2.6),
α ∧ k ` (1, −1) = k −1 (α) ∧ k `−1 (1, −1) . (5.3.52)

We thus obtain,
iη −1 (α)∧ 1 k `+2 (1,−1) −1 (α+k ` (1,−1))

F ϕm = c e− 2h k 2 Tk ϕ`+2
iη −1
− 2h k (α+k` (1,−1))∧ 12 k`+2 (1,−1) k−1 (α+k` (1,−1))
= ce T ϕ`+2
as desired.
For the third relation, we first remark that
1
z(mα,` ) = z(α) + z(k ` (1, −1))
2
1
= z(α) + k 3−` (1, −1)
2
1
= z(α) + k 3−` (1, −1) + k −` (1, −1) .
2
So we have to prove that
iη 3−` (1,−1))∧ 1 k −` (1,−1) 3−` (1,−1)

Sϕm = d e− 2h (z(α)+k 2 T z(α)+k ϕ−`
iη
− 2h z(α)∧ 12 k−` (1,−1) z(α)+k3−` (1,−1)
= de T ϕ−`
for some d ∈ S 1 .
Let N = N(0,1),5 ∈ Γ and ϕn defined by (5.3.42). Hypothesis (5.2.30) and the fact
that χ is real, imply that sVm1 = Vn . Hence, after Lemma 5.9, there is a complex
number d, such that |d| = 1 and
Sϕ1 = d ϕn . (5.3.53)
Hence,
iη 1 5 2 (1,−1)
Sϕ1 = d e− 2h (0,1)∧ 2 k (1,−1)
T (0,1) F 2 ϕ1 = d T k F 2 ϕ1 , (5.3.54)
which, together with (5.3.35), (5.3.37) and (5.3.22), gives
iη 1 `
Sϕm = e 2h α∧ 2 k (1,−1)
T z(α) F `−1 Sϕ1
iη 1 ` 2 (1,−1)
= d e 2h α∧ 2 k (1,−1)
T z(α) F `−1 T k F 2 ϕ1 (5.3.55)
iη
α∧ 12 k` (1,−1) k3−` (1,−1)
= de 2h T z(α) T F 1−(−`) ϕ1
iη iη
α∧ 12 k` (1,−1) z(α)∧k3−` (1,−1) 3−` (1,−1)
= de 2h e 2h T z(α)+k ϕ−` .
We remark that −k 3−` (1, −1) = k −` (1, −1) and that after (5.2.22),
α ∧ k ` (1, −1) = −z(α) ∧ z(k ` (1, −1)) = z(α) ∧ k −` (1, −1) . (5.3.56)
We thus obtain,
iη 1 −` 3−` (1,−1)
Sϕm = d e− 2h z(α)∧ 2 k (1,−1)
T z(α)+k ϕ−`
iη
− 2h (z(α)+k3−` (1,−1))∧ 12 k−` (1,−1) 3−` (1,−1)
= de T z(α)+k ϕ−`
as desired.
Remark 5.19. In the case of a constant magnetic field, using (5.3.40) we verify
that
c = 1 and d = 1 .
(Step 3 ) We may now state Carlsson’s result. Let Σ be the spectral space associated
with I(h) and Π the orthogonal projection over Σ. We define the projections
vm = Π ϕm , M ∈ Γ.
We denote

dV (M, K1 ) + dV (K1 , K2 ) + · · · + dV (K`−1
˜ , N) ;
d(`)
m,n = inf
˜
`≥` M 6= K1 6= K2 6= · · · =
6 K`−1
˜ 6= N
where the points Kj belong to the kagome lattice Γ. By estimates (5.3.43) and
(5.3.44), for every ε > 0 we can choose δ > 0 in the definition of χ such that
1−ε
| hvm , vn i | ≤ d(1)
m,n
for h ∈ (0, h(ε)). We denote D the matrix hvn , vm i and we define the functions
X
em = vn (D−1/2 )m,n .
M ∈Γ
Carlsson’s theorem reads in our case as follows.

Theorem 5.20. There exists h0 > 0 such that for h ∈ (0, h0 ), the functions em ,
M ∈ Γ, form an orthonormal basis of Σ. The matrix of Ph,A,V in this basis is given
by
λ(h) id + w
where λ(h) is the first eigenvalue of Pm1 and
wm,n = h(Ph,A,V − λ(h)) en , em i . (5.3.57)

Moreover,
w = w̃ + D̃ε
where
1
w̃m,n =hϕn , rm i + hrn , ϕm i (5.3.58)
2
and for every ε > 0 we can choose δ in the definition of χ in Step 1 such that
1−ε
(D̃ε )m,n ≤ d(2)
m,n
for h ∈ (0, h(ε)).
(Step 4 ) We finally prove that the orthonormalization process preserves the action
of the magnetic operators.
Proposition 5.21. For every M ∈ Γ and β ∈ Z2 we have
iη
T β em = e 2h β∧m em+β (5.3.59)
F em = c ek−1 (m) (5.3.60)
Sem = d ez(m) , (5.3.61)
where c ∈ {−1, 1} and d ∈ S 1 are defined in Proposition 5.18.
Proof. After Lemma 5.11, T β commutes with Ph,A,V , so also with Π using the
functional calculus of Ph,A,V . Using (5.3.46) and the definition of vn we get
X
T β em = T β vn (D−1/2 )m,n
n
X iη
= e 2h β∧n vβ+n (D−1/2 )m,n (5.3.62)
n
iη X iη iη
= e 2h β∧m vβ+n e− 2h β∧m (D−1/2 )m,n e 2h β∧n .
n
Now, since T β is unitary, using (5.3.46) again, we find

iη iη
D̂m,n := hvβ+n , vβ+m i = e− 2h β∧m Dm,n e 2h β∧n .
iη
Considering the diagonal matrix Am,m = e− 2h β∧m , we note that D̂ = ADA−1 and
that
iη iη
(ADA−1 )−1/2 = e− 2h β∧m (D−1/2 )m,n e 2h β∧n .

m,n
So (5.3.62) becomes
iη X
T β em = e 2h β∧m vβ+n (D̂−1/2 )m,n .
n
The sum in the right hand side is actually the vector corresponding to m + β in the
orthonormalization of {vn } which yields (5.3.59).
Similarly, using (5.3.47), we find

X
F em = c vk−1 (n) (D−1/2 )m,n , (5.3.63)
n
and since the magnetic rotation is a unitary operator, we have D = D̃, where

D̃m,n = vk−1 (n) , vk−1 (m) . (5.3.64)
Hence,
X
F em = c vk−1 (n) (D̃−1/2 )m,n .
n
The right hand side of the previous equality is c times the vector corresponding to
k −1 (m) in the orthonormalization of {vn }, so we get relation (5.3.60). Using (5.3.48),
an analogous computation gives (5.3.61).
5.4 The interaction matrix

5.4.1 Introduction
In this section we perform the reduction of the operator Ph,A,V to different discrete
models acting on `2 (Z2 ), `2 (Z) or C3q , q ∈ N∗ . We also present some ideas about the
pseudo-differential operator related to the low lying spectrum of Ph,A,V .
We first give the properties of the matrix w introduced in Theorem 5.20. Then, under
some additional hypotheses, we construct a new operator Wγ by keeping the main
terms of w and neglecting the terms which do not correspond to nearest neighbors for
the Agmon distance. Here γ is the normalized magnetic flux trough a fundamental
domain of the one of the triangular lattices spanned by 2ν1 and 2ν2 and translated
by ν` , ` ∈ {1, 3, 5}. Next, we express Wγ using discrete translations. Finally, in the
case of a rational value of the normalized flux, we reduce the operator to a family
of hermitian matrices. We end by presenting the picture of the butterfly associated
5.4. THE INTERACTION MATRIX 169
with the low lying spectrum of Ph,A,V . We refer to the work of Bellissard (see for
example [29]) about C∗ -algebras for more details about the relations between the
different representations of Wγ .
5.4.2 The nearest neighbors interaction matrix

The matrix w in the basis {em }M ∈Γ of the restriction of Ph,A,V to Σ is given in
Theorem 5.20 (see (5.3.57)). Using Lemmas 5.9 and 5.11 together with properties
(5.3.59), (5.3.60) and (5.3.61), we find
iη
wm,n = e− 2h (n−m)∧γ w(m+γ),(n+γ) (5.4.1)
wm,n = wk(m),k(n) (5.4.2)
wm,n = wz(m),z(n) . (5.4.3)
By (5.4.1) we have
iη iη
e− 2h n∧m wm,n = e− 2h (n+γ)∧(m+γ) wm+γ,n+γ . (5.4.4)
Hence, the left hand size of the previous equality only depends on the difference
n − m. In order to write in a simpler way the matrix w, for `, j ∈ {1, 3, 5}, ` 6= j,
we define the functions g`,j : Z2 → C by
iη
g`,j (β − α) = e− 2h mβ,j ∧mα,` wmα,` ,mβ,j . (5.4.5)
The family of functions g`,j inherits the properties of the hermitian matrix w. We
first have
g`,j (γ) = gj,` (−γ) , γ ∈ Z2 . (5.4.6)

Also, denoting by ν̃` the coordinates of the vector ν` in the basis B (see the Section
5.2.1), relation (5.4.2) together with (5.2.23) give
g`,j (γ) = g`+2,j+2 (k −1 (γ) + ν̃j+1 − ν̃`+1 ) . (5.4.7)

Finally, relation (5.4.3) implies
g`,j (γ) = g−`,−j (z(γ) + ν̃j−3 − ν̃`−3 ) . (5.4.8)

Nearest neighbors
For any α ∈ Z2 , we want to identify in the matrix w, the main terms corresponding
to the interactions between the nearest wells for the Agmon distance to the triple
{Mα,1 , Mα,3 , Mα,5 }. The others terms being shown to be relatively small and there-
fore neglected. The determination of the nearest wells may depend on the particular
choice of the electric potential V . Here we assume (see [92] for more details) :
Hypothesis 5.22.
A. The nearest neighbors for the Agmon distance are the same of those for the Eu-
clidean distance.
B. Between two nearest neighbors M, N ∈ Γ there exists an unique minimal geodesic

γM,N for the Agmon metric.
C. This geodesic γM,N coincides with the Euclidean one that is the segment between
M and N .
D. The geodesic γM,N in non degenerate in the sense that there is a point 5 x̃ ∈
γM,N \ {M, N } such that the function x 7→ dV (x, M ) + dV (x, N ) − dV (M, N )
restricted to a transverse line to γM,N at x̃ has a non degenerate local minimum
at x̃.
Depending on the result we could assume all or some of the previous hypothesis. To
simplify the analysis, here we assume all the assumptions.
Denoting by N (M ) the set of nearest neighbors of M for the Agmon distance,

Hypothesis 5.22 A yields (see Figure 5.7) :
N (Mα,1 ) = {M(α+ν̃1 ),3 , M(α−ν̃3 ),3 , M(α+ν̃1 ),5 , M(α−ν̃5 ),5 } ,
N (Mα,3 ) = {M(α−ν̃1 ),1 , M(α+ν̃3 ),1 , M(α−ν̃5 ),5 , M(α+ν̃3 ),5 } , (5.4.9)
N (Mα,5 ) = {M(α−ν̃1 ),1 , M(α+ν̃5 ),1 , M(α+ν̃5 ),3 , M(α−ν̃3 ),3 } .
Structure of the main term
Neglecting in w the terms which do not not correspond to the nearest neighbors
listed in (5.4.9), we obtain a new operator Wγ acting on elements v = (v 1 , v 3 , v 5 ) ∈
`2 (Z2 ) × `2 (Z2 ) × `2 (Z2 ) by
γ γ
(Wγ v)1n,m = ei 8 (2n+1) g1,3 (ν̃1 ) vn+1,m
3
+ e−i 8 (2n+1) g1,3 (−ν̃3 ) vn+1,m−1
3
γ γ
+ e−i 8 (2n+2m+1) g1,5 (ν̃1 ) vn+1,m
5
+ ei 8 (2n+2m+1) g1,5 (−ν̃5 ) vn,m+1
5
γ γ
(Wγ v)3n,m = e−i 8 (2n−1) g3,1 (−ν̃1 ) vn−1,m
1
+ ei 8 (2n−1) g3,1 (ν̃3 ) vn−1,m+1
1
γ γ
+ e−i 8 (2m+1) g3,5 (−ν̃5 ) vn,m+1
5
+ ei 8 (2m+1) g3,5 (ν̃3 ) vn−1,m+1
5
γ γ
(Wγ v)5n,m = ei 8 (2n+2m−1) g5,1 (−ν̃1 ) vn−1,m
1
+ e−i 8 (2n+2m−1) g5,1 (ν̃5 ) vn,m−1
1
γ γ
+ ei 8 (2m−1) g5,3 (ν̃5 ) vn,m−1
3
+ e−i 8 (2m−1) g5,3 (−ν̃3 ) vn+1,m−1
3
.
Here γ = γ(h, η) is the normalized flux of the magnetic field trough a fundamental
domain of one of the triangular lattice Γ` , ` ∈ {1, 3, 5}, given by
η
γ=− . (5.4.10)
h
5. Actually this condition does not depend on the choice of the point x̃ (see [92]).
Remark 5.23. In the case of a constant magnetic field, the normalized flux of
the magnetic field trough a triangle having a point of the triangular lattice Γ` , ` ∈
{1, 3, 5}, in each of its vertex (see Figure 5.7) is
γ
γ4 = .
8
This reduction is justified by the Agmon estimates in (5.3.43) and (5.3.44), together
with the structure of the matrix w given in Theorem 5.20.
We now give a first simplification of the operator Wγ using the properties of the
functions g`,j . Using relation (5.4.7) twice we get
g1,3 (ν̃1 ) = g3,5 (ν̃3 ) = g5,1 (ν̃5 ) ,

g1,3 (−ν̃3 ) = g3,5 (−ν̃5 ) = g5,1 (−ν̃1 ) ,
(5.4.11)
g1,5 (−ν̃5 ) = g3,1 (−ν̃1 ) = g5,3 (−ν̃3 ) ,
g1,5 (ν̃1 ) = g3,1 (ν̃3 ) = g5,3 (ν̃5 ) .
Relation (5.4.6) gives
g1,3 (ν̃1 ) = g3,1 (−ν̃1 ) ,

(5.4.12)
g1,5 (ν̃1 ) = g5,1 (−ν̃1 ) .
Moreover, using (5.4.7) again we obtain
g1,3 (ν̃1 ) = g3,5 (−ν̃5 ) . (5.4.13)

Relations (5.4.11)-(5.4.13) imply that all the coefficients g`j (·) in the operator Wγ
can be written in terms of g1,3 (ν1 ). We introduce
ρ = |g1,3 (ν̃1 )| > 0 and β = π + Arg(g1,3 (ν̃1 )) . (5.4.14)

We remark that β and ρ depend on h and the magnetic flux η, and that ρ is related
to the magnitude of the tunneling effect and to the length of the band containing
to the low lying spectrum.
After division by −ρ our system takes the form

i γ8 (2n+1) −i γ8 (2n+1)
(Wγ v)1n,m = e iβ
e 3
vn+1,m +e 3
vn+1,m−1
γ γ

+ e−iβ e−i 8 (2n+2m+1) vn+1,m
5
+ ei 8 (2n+2m+1) vn,m+1
5
γ γ

(Wγ v)3n,m = e−iβ e−i 8 (2n−1) vn−1,m
1
+ ei 8 (2n−1) vn−1,m+1
1
γ
iβ −i 8 (2m+1) 5 i γ8 (2m+1) 5
+ e e vn,m+1 + e vn−1,m+1
γ γ

(Wγ v)5n,m = eiβ ei 8 (2n+2m−1) vn−1,m
1
+ e−i 8 (2n+2m−1) vn,m−1
1
γ γ

+ e−iβ ei 8 (2m−1) vn,m−1
3
+ e−i 8 (2m−1) vn+1,m−1
3
,
or equivalently,
γ γ

(Wγ v)1n,m = eiβ ei 8 (2n+1) vn+1,m
3
+ e−i 8 (2n+1) vn+1,m−13
γ
−iβ −i 8 (2n+2m+1) 5 i γ8 (2n+2m+1) 5
+ e e vn+1,m + e vn,m+1
γ γ

(Wγ v)3n+1,m = e−iβ e−i 8 (2n+1) vn,m
1
+ ei 8 (2n+1) vn,m+1 1
γ
iβ −i 8 (2m+1) 5 i γ8 (2m+1) 5
+ e e vn+1,m+1 + e vn,m+1
γ γ

(Wγ v)5n,m+1 = eiβ ei 8 (2n+2m+1) vn−1,m+1
1
+ e−i 8 (2n+2m+1) vn,m 1
γ γ

+ e−iβ ei 8 (2m+1) vn,m
3
+ e−i 8 (2m+1) vn+1,m 3
.
In this way, the problem only depends on the two parameters γ and β representing
the “past” of the problem.
5.4.3 The representation using discrete translations

First representation on [`2 (Z2 )]3
Our goal now is to write Wγ in a compact way using the parameters γ, β and two
discrete translations τ̂1 , τ̂2 acting on `2 (Z2 ) by
γ γ
(τ̂1 v)n,m = ei 2 m vn−1,m and (τ̂2 v)n,m = e−i 2 n vn,m−1 .
These discrete translations satisfy
τ̂1 τ̂2 = eiγ τ̂2 τ̂1 . (5.4.15)
1 1 3 3 5 5
Introducing ṽn,m = vn,m , ṽn,m = vn+1,m and ṽn,m = vn,m+1 we find

i γ8 (2n+1) −i γ8 (2n+1)
(Wγ ṽ)1n,m = e iβ
e 3
ṽn,m +e 3
ṽn,m−1
γ γ

+ e−iβ e−i 8 (2n+2m+1) ṽn+1,m−1
5
+ ei 8 (2n+2m+1) ṽn,m
5
γ
−iβ −i 8 (2n+1) 1 i γ8 (2n+1) 1
(Wγ ṽ)3n,m = e e ṽn,m + e ṽn,m+1
γ γ

+ eiβ e−i 8 (2m+1) ṽn+1,m
5
+ ei 8 (2m+1) ṽn,m 5
γ γ

(Wγ ṽ)5n,m = eiβ ei 8 (2n+2m+1) ṽn−1,m+1
1
+ e−i 8 (2n+2m+1) ṽn,m
1
γ
−iβ i 8 (2m+1) 3 −i γ8 (2m+1) 3
+ e e ṽn−1,m + e ṽn,m ,
or equivalently,
γ γ

(Wγ ṽ)1n,m = eiβ ei 8 (2n+1) ṽn,m
3
+ e−i 8 (2n+1) ṽn,m−1
3
γ γ

+ e−iβ e−i 8 (2n+2m+1) ṽn+1,m−1
5
+ ei 8 (2n+2m+1) ṽn,m
5

i γ8 (2n+1) 3 −iβ 1 i γ8 (4n+2) 1
e (Wγ ṽ)n,m = e ṽn,m + e ṽn,m+1
γ γ

+ eiβ ei 8 (2n−2m) ṽn+1,m
5
+ ei 8 (2n+2m+2) ṽn,m
5
γ
i γ8 (2n+2m+1) 5 iβ i 8 (4n+4m+2) 1 1
e (Wγ ṽ)n,m = e e ṽn−1,m+1 + ṽn,m
γ γ

+ e−iβ ei 8 (2n+4m+2) ṽn−1,m3
+ ei 8 2n ṽn,m
3
.
γ γ
1
Then, introducing v̂n,m 1
= ṽn,m 3
, v̂n,m = ei 8 (2n+1) ṽn,m
3 5
and v̂n,m = ei 8 (2n+2m+1) ṽn,m
5
, we
obtain

−i γ8 (4n+2)
(Wγ v̂)1n,m = e iβ 3
v̂n,m
+e 3
v̂n,m−1
γ
+ e−iβ e−i 8 (4n+4m+2) v̂n+1,m−1
5 5
+ v̂n,m

−iβ i γ8 (4n+2) 1
(Wγ v̂)3n,m = e 1
v̂n,m + e v̂n,m+1
γ γ

+ eiβ e−i 8 (4m+3) v̂n+1,m
5
+ ei 8 v̂n,m
5
γ
(Wγ v̂)5n,m = eiβ ei 8 (4n+4m+2) v̂n−1,m+1
1 1
+ v̂n,m
γ
−iβ i 8 (4m+3) 3 −i γ8 3
+ e e v̂n−1,m + e v̂n,m .
We have then found for Wγ the following compact form
γ γ
eiβ (1 + e−i 4 τ̂2 ) e−iβ (1 + e−i 4 τ̂1∗ τ̂2 )
 
0
γ γ 3γ
Wγ =  e−iβ (1 + ei 4 τ̂2∗ ) 0 eiβ (ei 8 + e−i 8 τ̂1∗ )  . (5.4.16)
γ γ 3γ
eiβ (1 + ei 4 τ̂2∗ τ̂1 ) e−iβ (e−i 8 + ei 8 τ̂1 ) 0
Computation of the tunneling effect
In order to determine the spectrum of Wγ , we need at least to estimate a value for

β in function of h and the magnetic flux η. In the case of a square, triangular or
hexagonal lattice (see [94] and [111]), using the commutation relations between the
magnetic operators it can be proved that β = 0. In the case of the kagome lattice,
the geometry of the problem is more complicate and we are not able to do this.
However, under additional conditions about the electric and magnetic potentials,
we will approximate β in the semi-classical limit.
The parameter β is related to the interaction between the wells of the kagome lattice.
As in [94] and [111], we implement here the results of [93] about the tunneling
effect to estimate β and at the same time ρ. This justify the approximation of the
original problem by Wγ . We give the estimations in the case of a constant and weak
magnetic field B = hB0 , with B0 a positive constant and when the gauge for the
vector potential is
hB0
A(x1 , x2 ) =
(−x2 , x1 ) . (5.4.17)
2
In this case, the value of the normalized flux through a cell of periodicity of the
lattice spanned by 2ν1 and 2ν2 (see equality (5.3.27) in the proof of Proposition
5.14) is
η √
= B0 (2ν1 ) ∧ (2ν2 ) = 2 3 B0 . (5.4.18)
h
Moreover, as we have seen through this chapter, in this case we can compute expli-
citly the value of the phases of the magnetic translations and rotation. Indeed, Re-
marks 5.13 and 5.15, together with (5.4.18), give that for any α ∈ Z2 and ϕ ∈ L2 (R2 )
B0
(T α ϕ)(x) = e−i x∧(2α1 ν1 +2α2 ν2 )
ϕ τ −α (x)

2 and (F ϕ)(x) = ϕ(r(x)) . (5.4.19)
We recall that the definition of β is related to the interaction between the wells
M(0,0),1 and M(1,0),3 (see (5.4.14)) by
g1,3 (ν̃1 ) = ρ ei (π+β) , (5.4.20)

where the coefficient g1,3 (ν̃1 ) is defined in (5.4.5).
Proposition 5.24. Let ρ and β be given by (5.4.20) and the magnetic vector po-
tential A by (5.4.17). Under the Hypothesis 5.22, there is b0 > 0 and h0 > 0 such
that for h ∈ (0, h0 )
dV (M(1,0),3 ,M(0,0),1 )
ρ = h /2 b0 e−
1
h (1 + O(h)) ,
and
β = O(h) . (5.4.21)
From now on we suppose that the parameter β appearing in Wγ satisfies

(5.4.21).
Considering (5.4.16), we then derive that
Wγ = W̃γ + O(h) ,
where γ γ
1 + e−i 4 τ̂2 1 + e−i 4 τ̂1∗ τ̂2
 
0
γ γ 3γ
W̃γ =  1 + ei 4 τ̂2∗ 0 ei 8 + e−i 8 τ̂1∗  (5.4.22)
γ γ 3γ
1 + ei 4 τ̂2∗ τ̂1 e−i 8 + ei 8 τ̂1 0
and the parameter γ (see 5.4.10) is given by
√
γ = −2 3 B0 . (5.4.23)
Proof of Proposition 5.24. By (5.4.5) and (5.4.18) we have
η
g1,3 (ν̃1 ) = e−i 2h m(1,0),3 ∧m(0,0),1 wm(0,0),1 ,m(1,0),3
√
3 B0 12 (ν̃1 +ν̃2 )∧ 21 ν̃1
= e−i wm(0,0),1 ,m(1,0),3
√
3 B0
= ei 4 wm(0,0),1 ,m(1,0),3 . (5.4.24)
The results in Section 3 of [93] give an asymptotic estimate for wm(0,0),1 ,m(1,0),3 . In
order to apply the results there, we first need to verify that the values of the functions
ϕm(0,0),1 and ϕm(1,0),3 at the bottom of their respective wells are real. The value of
ϕm(0,0),1 (M(0,0),1 ) has been chosen real in (5.3.40). The definition of ϕm in (5.3.42),
together with (5.4.18) and (5.4.19), give
η 1
ϕm(1,0),3 (x) = e−i 2h ν̃1 ∧ 2 ν̃3 T 1 F −2 ϕ1 (x)

√
3 B0 B0
−i
e−i F −2 ϕ1 (x − 2ν1 )
x∧2ν1

= e 2 2
√
3 B0
= e−i e−iB0 x∧ν1 ϕ1 r−2 (x − 2ν1 ) .

2
√
Since M(1,0),3 = 2ν1 + ν3 we find M(1,0),3 ∧ ν1 = − 3/2, so
ϕm(1,0),3 M(1,0),3 = ϕ1 (r−2 (ν3 )) = ϕ1 (ν1 ) = ϕm(0,0),1 (M(0,0),1 ) .

Thus, ϕm(1,0),3 (M(1,0),3 ) is also real.
The results in [93] are given for a magnetic potential At = tA, with the condition
(see (2.40) there) |t| = O(h1/2 (−Log h)1/2 ). Our setting satisfies this requirement with
t = h. By Proposition 3.12, Remark 3.17 and Lemma 3.15 in [93] 6 and assuming
Hypothesis 5.22 we get
S(h)
wm(0,0),1 ,m(1,0),3 = h /2 b(h) e−
1
h , (5.4.25)
where
|b(h)| = b0 + O(h) , (5.4.26)

Arg(b(h)) = π + O(h) (5.4.27)
and
Re(S(h)) = dV (M(0,0),1 , M(1,0),3 ) + O(h2 ) , (5.4.28)

Im(S(h)) = Circ(A, γ) + O(h3 ) . (5.4.29)
Here before γ : [0, 1] → [M(0,0),1 , M(1,0),3 ] is the unique minimal geodesic between
M(0,0),1 and M(1,0),3 and
t
6. Formula (3.26) in [93] has unfortunately disappeared in the printing and reads : Wjk =
1/2 t −Sjk /h
h bjk (h)e .
Z
Circ(A, γ) = ωA .
γ
Considering (5.4.17) we obtain

√
3 B0
Circ(A, γ) = h. (5.4.30)
4
The result comes then putting together (5.4.20)-(5.4.30).
5.4.4 Representation using pseudo differential operators

We recall (see (5.1.2)) that the Harper model was
τ1 + τ1∗ + t(τ2 + τ2∗ ) ,

which corresponds to the Weyl quantification of the symbol
2 cos ξ + 2t cos x .
Here we consider the operator W̃γ defined in (5.4.22) and make the correspondence
τ̂1 = eix and τ̂2 = eiγDx .

As in [94] or [111], it can be shown that W̃γ is isospectral with the pseudo-differential
operator given by
γ γ
1 + e−i 4 eiγDx 1 + e−i 4 e−ix eiγDx
 
0
 1 + ei γ4 e−iγDx 0
γ 3γ
ei 8 + e−i 8 e−ix  . (5.4.31)
γ γ 3γ
1 + ei 4 e−iγDx eix e−i 4 + ei 8 eix 0
The γ-principal symbol is
1 + eiξ 1 + e−ix+iξ
 
0
 1 + e−iξ 0 1 + e−ix  . (5.4.32)
−iξ+ix ix
1+e 1+e 0
The characteristic polynomial is
∆(λ, x, ξ) = −λ3 + 2λ(3 + cos ξ + cos(ξ − x) + cos x) (5.4.33)

+4(1 + cos ξ + cos x + cos(ξ − x)) . (5.4.34)
Observing that λ = −2 is a root we get
∆(λ, x, ξ) = −(λ + 2) λ2 − 2λ − 2(1 + cos ξ + cos(ξ − x) + cos x))

or
∆(λ, x, ξ) = −(λ + 2) (λ − 1)2 − (3 + 2 cos ξ + 2 cos(ξ − x) + 2 cos x) . (5.4.35)

This lead to three roots :

p
λ1 = −2 , λ± (x, ξ) = 1 ± 3 + 2 cos ξ + 2 cos(ξ − x) + 2 cos x .
The range is [−2, 4] as was found in the numerical computations of Hou (see [99])
and in our numerical computations (see the bottom line in Figure 5.13 and (5.4.47)).
Looking at the multiplicity, we get
∂λ ∆(λ, x, ξ) = −3λ2 + 2(3 + cos ξ + cos(ξ − x) + cos x) .

The only case where there is multiplicity 2 is for λ = −2 (which seems to be a new
phenomenon which was not met in the butterfly analysis in [97]) and for λ = 1. This
corresponds to the points (x, ξ) such that
3 + 2 cos ξ + 2 cos(ξ − x) + 2 cos x = 0 ,

so we obtain (x, ξ) = (± 2π
3
, ∓ 2π
3
).
Hence we get three bands : one reduced to a point and touching a band [−2, 1], the
last one being [1, 4] with touching at 1.
Remark 5.25. The expression 2 cos ξ + 2 cos(ξ − x) + 2 cos x appearing in (5.4.35)

is the principal symbol in the case of the triangular lattice (see (5.1.11)).
Remark 5.26. The semi-classical study as γ tends to 0 or to a rational number is
an open problem.
Second representation on [`2 (Z)]3
As in the case of the square, triangular and hexagonal lattices (see (5.2)), we will
perform a partial Floquet theory with respect to the second variable. In order to do
this, we first write W̃γ using to different discrete translations τ1 , τ2 acting on `2 (Z2 )
by
(τ1 v)n,m = vn−1,m and (τ2 v)n,m = e−iγn vn,m−1 .

As in (5.4.15) we find that
τ1 τ2 = eiγ τ2 τ1 . (5.4.36)
i γ2 nm
Moreover, there exists an unitary operator U on `2 (Z2 ), given by (Uv)n,m = e vn,m ,
such that
Uτ1 U ∗ = τ̂1 and Uτ2 U ∗ = τ̂2 .

The operator W̃γ defined in (5.4.22) is then unitary equivalent with
γ γ
1 + e−i 4 τ2 1 + e−i 4 τ1∗ τ2
 
0
γ γ 3γ
W̌γ =  1 + ei 4 τ2∗ 0 ei 8 + e−i 8 τ1∗  . (5.4.37)
γ γ 3γ
1 + ei 4 τ2∗ τ1 e−i 8 + ei 8 τ1 0
Now, since the translations in the second variable commute with τ1 and τ2 , we can
perform the partial Floquet theory. We have
[
σ W̃γ = σ (Wγ,θ2 )
θ2 ∈[0,1[
where
1 1
1 + e−iγ (n+ 4 )+i2πθ2 1 + e−iγ e−iγ (n+ 4 )+i2πθ2 τ1∗
 
0
1
Wγ,θ2 =  1 + eiγ (n+ 4 )−i2πθ2
γ 3γ
ei 8 + e−i 8 τ1∗ ,
 
0
1
1 + eiγ (n+ 4 )−i2πθ2 τ1
γ 3γ
e−i 8 + ei 8 τ1 0
(5.4.38)
and
j
D(Wγ,θ2 ) = (v 1 , v 3 , v 5 ) ∈ `2 (Z2 ) × `2 (Z2 ) × `2 (Z2 ) ; vn,m−1 j
= ei2πθ2 vn,m

, j ∈ {1, 3, 5} .
Third discrete representation for a rational value of the flux
As we saw in the introduction of this chapter, we may perform a new Floquet

decomposition when the value of the normalized flux is equal to 2π times a rational
number, that is,
p
γ = 2π with p, q relatively prime . (5.4.39)
q
Indeed, in this case τ1q commutes with τ1 and with the multiplication by e±iγn . Thus,
we may use a Floquet theory and obtain
[
σγ = σ (Wp,q,θ1 ,θ2 ) (5.4.40)
(θ1 ,θ2 )∈[0,1[×[0,1[
where

σγ = W̃γ (5.4.41)
and
j
D(Wp,q,θ1 ,θ2 ) = (v 1 , v 3 , v 5 ) ∈ `2 (Z) × `2 (Z) × `2 (Z) ; vn+q = ei2πθ1 q vnj , j ∈ {1, 3, 5} .

Since the sequences v ∈ D(Wp,q,θ1 ,θ2 ) are determined by v0j , · · · , vq−1

j
, j ∈ {1, 3, 5},
this leads to the study of the hermitian matrices of size 3q acting on
v = (v 1 , v 3 , v 5 ) = (v01 , · · · , vq−1
1
, v03 , · · · , vq−1
3
, v05 , · · · , vq−1
5
)
by  
13 15
 0q Wp,q,θ1 ,θ2
Wp,q,θ1 ,θ2 


 
 
 
 13 ∗ 35

Wp,q,θ1 ,θ2 =
 Wp,q,θ1 ,θ2 0q Wp,q,θ1 ,θ2 
,
 
 
 
 15 ∗ 35 ∗ 
 Wp,q,θ1 ,θ2 Wp,q,θ1 ,θ2 0q 
where
π p
13
Wp,q,θ1 ,θ2
= Iq + e−i 2 q ei2πθ2 Jp,q
∗
π p
15
Wp,q,θ1 ,θ2
= Iq + e−i 2 q ei2πθ2 Jp,q
∗
Kq
π p 3π p
35
Wp,q,θ1 ,θ2
= ei 4 q Iq + e−i 4 q Kq ,
with
Jp,q = diag(exp (2iπ(j − 1)p/q))

and

 1 if j = i + 1 (mod q) and i 6= q
iq2πθ1
(Kq )ij = e if j = i + 1 (mod q) and i = n .
0 if j 6= i + 1 (mod q)

To further simplify the expression of Wp,q,θ1 ,θ2 , we observe that there exists an unitary
matrix V = diag(exp (2iπθ1 (j − 1))) such that
VJp,q V ∗ = Jp,q and VKq V ∗ = e2iπθ1 K

where

1 if j = i + 1 (mod q)
Kij = .
0 if j 6= i + 1 (mod q)
Hence, Wp,q,θ1 ,θ2 is unitary equivalent with the matrix
 
 0q M13
p,q,θ1 ,θ2 M15
p,q,θ1 ,θ2 


 
 
 
 13 ∗ 35

Mp,q,θ1 ,θ2 =
 Mp,q,θ1 ,θ2 0q Mp,q,θ1 ,θ2 
 (5.4.42)
 
 
 
 15 ∗ 35 ∗ 
 Mp,q,θ1 ,θ2 Mp,q,θ1 ,θ2 0q 
where
π p
−i 2 q i2πθ2 ∗
M13
p,q,θ1 ,θ2 = Iq + e e Jp,q
π p
−i 2 q 2iπ(θ1 +θ2 ) ∗
M15
p,q,θ1 ,θ2 = Iq + e e Jp,q K (5.4.43)
i π4 pq −i 3π p
M35
p,q,θ1 ,θ2 = e Iq + e 4 q e2iπθ1 K .
As in (5.4.15) and (5.4.36) we remark that Jp,q and K satisfy

p
KJp,q = e2iπ q Jp,q K .
5.4.5 Hofstadter’s butterfly for the kagome lattice

In Figure 5.13 we present the butterfly corresponding to the kagome lattice. We
consider σγ defined in (5.4.41) for γ = p/q and we numerically diagonalize Mp,q,θ1 ,θ2
given in (5.4.42) for θ1 , θ2 ∈ [0, 1].
In Figure 5.12 we plot σγ for some rational values of γ ∈ [0, 2]. We notice that for
fixed γ = p/q the spectrum is composed of 3q bands.
Looking at the expression of the matrices M`j p,q,θ1 ,θ2 in (5.4.43), we notice that the
smallest positive integer k for which the operator is invariant by the transformation
γ 7→ γ + k is k = 8. This is also clear from the definition of W̃γ using the discrete
translations τ1 and τ2 or from the pseudo-differential operator given in (5.4.31).
Hence, we have the property :
σγ+8 = σγ (translation invariance) . (5.4.44)

We thus plot in the vertical axis of the Figure 5.13 the bands of the spectrum for
p
γ = 2π and p, q relatively prime with p < 8q ≤ 400 .
q
We observe that for every N ∈ Z, ∪γ∈[N,N +8] σγ is composed of 8 different parts.
Moreover,
1. If N ≡ 2 [8], then ∪γ∈[N,N +1] σγ is deduce from ∪γ∈[1,2] σγ after a rotation of
angle π.
2. If N ≡ 3 [8], then ∪γ∈[N,N +1] σγ is deduce from ∪γ∈[0,1] σγ after a rotation of

angle π.
3. If N ≡ 4 [8], then ∪γ∈[N,N +1] σγ is deduce from ∪γ∈[0,1] σγ after a reflexion of

axis e = 0.

axis e = 0.

axis γ = 3/2.

axis γ = 1/2.
We find the following properties for the spectrum :
σ−γ = σγ (reflexion with respect to the axis γ = 0) (5.4.45)

e ∈ σγ+4 ⇔ −e ∈ σγ (reflexion with respect to the axis e = 0) (5.4.46)
Property (5.4.46) follows from the definition of Wp,q,θ1 ,θ2 . Indeed, for (`, j) ∈ {(1, 3), (1, 5), (3, 5)}
we have
M`j `j
−p,q,θ1 ,θ2 = Mp,q,−θ1 ,−θ2 ,
which give
M−p,q,θ1 ,θ2 = [Mp,q,−θ1 ,−θ2 ]∗ .

Hence, after (5.4.40) we obtain (5.4.45).
We also notice that (5.4.44) together with (5.4.45) imply
σ8−γ = σγ (reflexion with respect to the axis γ = 4) . (5.4.48)

(a)
(b)
Figure 5.12 – Spectrum of Wγ for some rational values of (a) γ ∈ [1, 2] and (b)
γ ∈ [0, 1].
Figure 5.13
– Hofstadter’s
butterfly for
the kagome
lattice.
Bibliographie
[1] Abo-Shaeera, J. R., Raman, C., Vogels, J. M., and Ketterle, W.

Observation of vortex lattices in Bose-Einstein condensates. Science 292, 5516
(2001), 476–479.
[2] Abrikosov, A. A. Nobel Lecture : Type-II superconductors and the vortex
lattice. Rev. Mod. Phys. 76, 3 (2004), 975–979.
[3] Acerbi, E., and Bouchitté, G. A general class of phase transition models
with weighted interface energy. Ann. Inst. H. Poincaré AN 25 (2008), 1111–
1143.
[4] Aftalion, A. Vortices in Bose-Einstein Condensates, vol. 67 of Progress in
Nonlinear Differential Equations and Their Applications. Birkhäuser, 2006.
[5] Aftalion, A., Alama, S., and Bronsard, L. Giant vortex and the break-
down of strong pinning in a rotating Bose-Einstein condensate. Arch. Rational
Mech. Anal. 178, 2 (2005), 247–286.
[6] Aftalion, A., Blanc, X., and Dalibard, J. Vortex patterns in a fast
rotating Bose-Einstein condensate. Phys. Rev. A 71, 023611 (2005).
[7] Aftalion, A., Blanc, X., and Nier, F. Lowest landau level for Bose
Einstein condensates and Bargmann transform. J. Funct. Anal. 241 (2006),
661–702.
[8] Aftalion, A., Blanc, X., and Nier, F. Vortex distribution in the lowest
Landau level. Phys. Rev. A 73, 011601 (2006).
[9] Aftalion, A., Dalibard, J., and Josserand, C. Polycopié du cours
Equation de Schrödinger non linéaire : des condensats de Bose Einstein aux
supersolides, École Polytechnique, 2011. http://catalogue.polytechnique.
fr/site.php?id=244&fileid=1433.
[10] Aftalion, A., and Du, Q. Vortices in a rotating Bose-Einstein condensate :
Critical angular velocities and energy diagrams in the Thomas-Fermi regime.
Phys. Rev. A 64, 063603 (2001).
[11] Aftalion, A., and Helffer, B. On mathematical models for Bose-Einstein
condensates in optical lattices. Rev. Math. Phys. 21, 2 (2008), 229–278.
[12] Aftalion, A., Jerrard, R. L., and Royo-Letelier, J. Non-existence
of vortices in the small density region of a condensate. J. Funct. Anal. 260
(2011), 2387–2406.
[13] Aftalion, A., and Mason, P. Rotation of a Bose Einstein condensate held
under a toroidal trap. Phys. Rev. A 81, 023607 (2010).
185
186 BIBLIOGRAPHIE
[14] Aftalion, A., Mason, P., and Wei, J. Vortex-peak interaction and lattice
shape in rotating two-component Bose-Einstein condensates. Phys. Rev. A 85,
033614 (2012).
[15] Aftalion, A., and Nier, F. Adiabatic approximation for a two-level
atom in a light beam. To appear in Ann. Scien. Fac. Toulouse. Preprint
in http ://arxiv.org/abs/1111.3811. (2012).
[16] Alama, S., Bronsard, L., and Montero, J. A. Vortices for a rotating
toroidal Bose–Einstein condensate. Arch. Rational Mech. Anal. 187, 3 (2008),
481–522.
[17] Alberti, G. Variational models for phase transitions, an approach via Γ-
convergence. Calculus of Variations and Differential Equations, Springer,
Berlin (2000), 95–114.
[18] Alberti, G., Bouchitté, G., and Seppecher, P. Phase transition with
the line-tension effect. Arch. Rational Mech. Anal. 144, 1 (1998), 1–46.
[19] Alinhac, S., and Gérard, P. Opérateurs pseudo-différentiels et théorème
de Nash-Moser. EDP Sciences, 1991.
[20] Alt, H. W., Caffarelli, L. A., and Friedman, A. Variational problems
with two phases and their free boundaries. Trans. Amer. Math. Soc. 282, 2
(1984), 431–461.
[21] Ambrosio, L., Fusco, N., and Pallara, D. Functions of bounded va-
riation and free discontinuity problems. Oxford New York : Clarendon Press,
2000.
[22] Ambrosio, L., and Tortorelli, V. M. Approximation of functionals de-
pending on jumps by elliptic functionals via Γ-convergence. Comm. Pure Appl.
Math. 43, 8 (1990), 999–1036.
[23] Ambrosio, L., and Tortorelli, V. M. On the approximation of free
discontinuity problems. Bolletino U.M.I. 7, 6-B (1992), 105–123.
[24] Anderson, M. H., Ensher, J. R., Matthews, M. R., Wieman, C. E.,
and Cornell, E. A. Observation of Bose-Einstein condensation in a dilute
atomic vapor. Science 269, 5221 (1995), 198–201.
[25] Attouch, H. Variational convergence for functions and operators. Pitman
Advanced Publishing Program, 1984.
[26] Avila, A., and Jitomirskaya, S. The Ten Martini Problem. Ann. of Math.
170 (2009), 303–342.
[27] Baldo, S., Jerrard, R., Orlani, G., and Soner, H. Vortex density
models for superconductivity and superfluidity. Comm. Math. Phys. 318, 1
(2013), 131–171.
[28] Bellissard, J. C∗ -algebras in solid state physics-2d electrons in a uniform
magnetic field. Warwick conference on operator algebras (1987).
[29] Bellissard, J. Le Papillon de Hofstadter, d’après B. Helffer et J. Sjöstrand.
Séminaire Bourbaki, 44ème année, 1991-92, 745, (novembre 1991). Astérisque
206 (1992), 7–40.
[30] Bellissard, J., and Simon, B. Cantor spectrum for the almost Mathieu
equation. J. Funct. Anal. 48, 3 (1982), 408–419.
BIBLIOGRAPHIE 187
[31] Berestycki, H., Lin, T.-C., Wei, J., and Zhao, C. On phase-separation
model : Asymptotics and qualitative properties. Arch. Rational Mech. Anal.
208 (2013), 163–200.
[32] Berestycki, H., Nirenberg, L., and Varadhan, S. R. S. The princi-
pal eigenvalue and maximum principle for second-order elliptic operators in
general domains. Comm. Pure and Appl. Math. 47, 47-92 (1994).
[33] Berestycki, H., Terracini, S., Wang, K., and Wei, J. On entire so-
lutions of an elliptic system modeling phase separations. Preprint (2012), to
appear.
[34] Bethuel, F., Brezis, H., and Hélein, F. Ginzburg-Landau vortices,
vol. 13. Birkhäuser, 1994.
[35] Bloch, I., Dalibard, J., and Zwerger, W. Many-body physics with
ultracold gases. Rev. Mod. Phys. 80, 885 (2008), 885–964.
[36] Bonnaillie-Noël, V. Nodal domains and spectral minimal par-
titions. http://w3.bretagne.ens-cachan.fr/math/simulations/
MinimalPartitions.
[37] Bonnetier, E., and Chambolle, A. Computing the equilibrium confi-
guration of epitaxially strained crystalline films. SIAM J. Appl. Math. 62, 4
(2002), 1093–1121.
[38] Bonnetier, E., Chambolle, A., and Jouve, F. A variational model for
equilibrium shapes of hetero-epitaxial films. PICOF’02 : 2ème Colloque sur
les problèmes inverses, le contrôle et l’optimisation de formes, M. Jaoua, J.
Jaffré, and T. Ha Duong, eds., Carthage, Tunisia (2002), 235–240.
[39] Bose, S. N. Plancks Gesetz und Lichtquantenhypothese. Z. Phys 26 (1924),
178.
[40] Bouchitté, G. Singular perturbations of variational problems arising from
a two-phase transition model. Appl. Math. Optim. 21, 3 (1990), 289–314.
[41] Bourdin, B., Bucur, D., and Oudet, E. Optimal partitions for eigenva-
lues. SIAM J. Sci. Comput. 31, 6 (2010), 4100–4114.
[42] Braides, A. Approximation of free-discontinuity problems. Lecture Notes in
Mathematics, Vol. 1694. Springer, 1998.
[43] Brascamp, H. J., and Lieb, E. H. On extensions of the Brunn-Minkowski
and Prékopa-Leindler theorems, including inequalities for log concave func-
tions, and with an application to the diffusion equation. J. Funct. Anal. 22
(1976), 366–389.
[44] Brezis, H. Analyse Fonctionnelle. Masson, 1983.
[45] Brezis, H. Semilinear equations in Rn without conditions at infinity. Appl.
Math. Optim. 12 (1984), 271–282.
[46] Brothers, J., and Ziemer, W. Minimal rearrangements of Sobolev func-
tions. J. Reine Angew. Math. 384 (1988), 153–179.
[47] Bucur, D., Buttazzo, G., and Henrot, A. Existence results for some
optimal partition problems. Adv. Math. Sci. Appl. 8, 2 (1998), 571–579.
188 BIBLIOGRAPHIE
[48] Caffarelli, A., and Kenig, C. E. Gradient estimates for variable coeffi-
cient parabolic equations and singular perturbation problems. Amer. J. Math.
120, 2 (1998), 391–439.
[49] Caffarelli, L. A., Karakhanyan, A. L., and Lin, F.-H. The geometry
of solutions to a segregation problem for nondivergence systems. J. Fixed
Point Theory Appl. 5 (2009), 319–351.
[50] Caffarelli, L. A., and Lin, F.-H. An optimal partition problem for ei-
genvalues. J. Sci. Comput. 31, 1-2 (2007), 5–18.
[51] Caffarelli, L. A., and Lin, F.-H. Singularly perturbed elliptic systems
and multi-valued harmonic functions with free boundaries. J. Amer. Math.
Soc. 21, 3 (2008), 847–862.
[52] Cartier, P. Quantum mechanical commutation relations and θ functions.
Proc. Symp. Pure Math, Boulder (1965), 183–186.
[53] Chang, S.-M., Lin, C.-S., Lin, T.-C., and Lin, W.-W. Segregated nodal
domains of two-dimensional multispecies Bose-Einstein condensates. Physica
D : Nonlinear Phenomena 196, 3-4 (2004), 341–361.
[54] Chu, S. Nobel Lecture : The manipulation of neutral particles. Rev. Mod.
Phys. 70, 3 (1998), 685–706.
[55] Claro, F. H., and Wannier, G. H. Magnetic subband structure of elec-
trons in hexagonal lattices. Phys. Rev. B 19, 12 (1979), 6068–6074.
[56] Cohen-Tannoudji, C. N. Nobel Lecture : Manipulating atoms with pho-
tons. Rev. Mod. Phys. 70, 3 (1998), 707–719.
[57] Conti, M., Terracini, S., and Verzini, G. An optimal partition problem
related to nonlinear eigenvalue. J. Funct. Anal. 198, 160-196 (2003).
[58] Conti, M., Terracini, S., and Verzini, G. Asymptotic estimates for the
spatial segregation of competitive systems. Adv. Math. 195, 2 (2005), 524–560.
[59] Conti, M., Terracini, S., and Verzini, G. On a class of optimal partition
problem related to the Fučík spectrum and to the monotonicity formulae. Calc.
Var. Partial Differential Equations 22, 1 (2005), 45–72.
[60] Conti, M., Terracini, S., and Verzini, G. A variational problem for the
spatial segregation of reaction-diffusion systems. Indiana Univ. Math. J. 54,
3 (2005), 779–815.
[61] Cornell, E. A., and Wieman, C. E. Nobel Lecture : Bose-Einstein conden-
sation in a dilute gas, the first 70 years and some recent experiments. Rev.
Mod. Phys. 74, 3 (2002), 875–893.
[62] Correggi, M., Rougerie, N., and Yngvason, J. The transition to a
giant vortex phase in a fast rotating Bose-Einstein condensate. Comm. Math.
Phys. 303 (2011), 451–608.
[63] Cycon, H., Froese, R., Kirsch, W., and Simon, B. Schrödinger opera-
tors. Springer, 1987.
[64] Dalibard, J. Polycopié du cours Atomes Froids, master Concepts
Fondamentaux de la Physique, École Normale Supérieure, année 2006.
http://www.phys.ens.fr/~dalibard/Notes_de_cours/DEA_atomes_
froids_actuel.pdf.
BIBLIOGRAPHIE 189
[65] Dalibard, J., Gerbier, F., Juzeliūnas, G., and Öhberg, P. Collo-
quium : Artificial gauge potentials for neutral atoms. Rev. Mod. Phys. 83, 4
(2011), 1523–1543.
[66] Damski, B., Fehrmann, H., Everts, H.-U., Baranov, M., Santos, L.,
and Lewenstein, M. Quantum gases in trimerized kagomé lattices. Phys.
Rev. A 72 (2005).
[67] Dancer, E. N., Wei, J., and Weth, T. A priori bounds versus multiple
existence of positive solutions for a nonlinear Schrodinger system. Ann. Inst.
H. Poincaré Anal. Non Linéaire 27 (2012), 953–969.
[68] Davis, K. B., Mewes, M. O., Andrews, M. R., van Druten, N. J.,
Durfee, D. S., Kurn, D. M., and Ketterle, W. Bose-Einstein conden-
sation in a gas of sodium atoms. Phys. Rev. Lett. 72, 22 (1995), 3969–3973.
[69] Delannoy, G., Murdoch, S. G., Boyer, V., Josse, V., Bouyer, P.,
and Aspect, A. Understanding the production of dual Bose-Einstein conden-
sation with sympathetic cooling. Phys. Rev. A 63 (2001), 051602(R).
[70] Dimassi, M., and Sjöstrand, J. Spectral Asymptotics in the Semi-Classical
Limit. Cambridge University Press, 1999.
[71] Ehrhard, A. Symétrisation dans l’espace de Gauss. Math. Scand. 53, 2
(1983), 281–301.
[72] Ehrhard, A. Inégalités isopérimetriques et intégrales de Dirichlet gaus-
siennes. Ann. Sci. École Norm. Sup. (4) 17, 2 (1984), 317–332.
[73] Einstein, A. Quantentheorie des einatomigen idealen Gases. Sitzber. Kgl.
Preuss. Akad. Wiss. (1924), 3–14 and 261–267.
[74] Evans, L. C., and Gariepy, R. F. Measure Theory and Fine Properties of
Functions. CRC Press, 1992.
[75] Ferrari, G., Inguscio, M., Jastrzebski, W., Modugno, G., and
Roati, G. Collisional properties of ultracold K-Rb mixtures. Phys. Rev.
Lett. 89, 5 (2002).
[76] Gallo, C. Expansion of the energy of the ground state of the Gross–Pitaevskii
equation in the Thomas–Fermi limit. Preprint on http://arxiv.org/abs/
1205.2975, 2012.
[77] Gallo, C., and Pelinovsky, D. On the Thomas–Fermi ground state in a
harmonic potential. Asymptotic Analysis 73 (2011), 53–96.
[78] Geim, A. K. Nobel Lecture : Random walk to graphene. Rev. Mod. Phys.
83, 3 (2011), 851–862.
[79] Gerbier, F., and Dalibard, J. Gauge fields for ultracold atoms in optical
superlattices. New J. Phys. 12 (2010).
[80] Gilbarg, D., and Trudinger, N. S. Elliptic Partial Differential Equations
of Second Order. Springer, 1977.
[81] Giorgi, E. D., and Franzoni, T. Su un tipo di convergenza variazionale.
Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur. 58, 842-850 (1975).
[82] Giusti, E. Minimal Surfaces and Functions of Bounded Variation. Mono-
graphs in Mathematics. Birkhäuser Boston, 1984.
190 BIBLIOGRAPHIE
[83] Günter, K., Cheneau, M., Yefsah, T., Rath, S., and Dalibard, J.
Practical scheme for a light-induced gauge field in an atomic Bose gas. Phys.
Rev. A 79, 011604 (2009).
[84] Hall, D., Matthews, M., Wieman, C., and Cornell, E. Measurements
of relative phase in binary mixtures of Bose-Einstein condensates. Phys. Rev.
Lett. 81 (1998), 1543–1547.
[85] Hall, D. S., Matthews, M. R., Ensher, J. R., Wieman, C. E., and
Cornell, E. A. Dynamics of component separation in a binary mixture of
Bose-Einstein condensates. Phys. Rev. Lett. 81, 8 (1998), 1539–1542.
[86] Helffer, B. Introduction to semi-classical methods for the Schrödinger ope-
rator with magnetic fields. In Aspects théoriques et appliqués de quelques EDP
issues de la géométrie ou de la physique. Proceedings of the CIMPA School
held in Damas (Syrie). Séminaires et Congrès. SMF, 2009.
[87] Helffer, B. On spectral minimal partitions : a survey. Milan J. Math. 78,
575-590 (2010).
[88] Helffer, B. Spectral Theory and its Applications, vol. 139 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, 2013.
[89] Helffer, B., and Hoffmann-Ostenhof, T. Converse spectral problems
for nodal domains. Mosc. Math. J. 7, 1 (2007), 67–84.
[90] Helffer, B., Hoffmann-Ostenhof, T., and Terracini, S. Nodal do-
mains and spectral minimal partitions. Ann. Inst. H. Poincaré Anal. Non
Linéaire 26, 1 (2009), 101–138.
[91] Helffer, B., Hoffmann-Ostenhof, T., and Terracini, S. On spectral
minimal partitions : the case of the sphere. Around the research of Vladimir
Mazya III 12 (2009), 153–178.
[92] Helffer, B., and Sjöstrand, J. Multiple wells in the semi-classical limit
I. Commun. in PDE 9, 4 (1984), 337–408.
[93] Helffer, B., and Sjöstrand, J. Effet tunnel pour l’équation de Schrödin-
ger avec champ magnétique. Ann. Scuola Norm. Sup. Pisa Vol XIV, 4 (1987),
625–657.
[94] Helffer, B., and Sjöstrand, J. Analyse semi-classique pour l’équation de
Harper (avec application à l’étude de Schrödinger avec champ magnétique).
Mémoire de la SMF No 34 ; Tome 116, Fasc. 4 (1988).
[95] Helffer, B., and Sjöstrand, J. Equation de Schrödinger avec champ
magnétique et équation de Harper, partie i champ magnétique fort, partie ii
champ magnétique faible, l’approximation de peierls. Lecture notes in Physics
(editors A. Jensen et H. Holden), 345 (1988), 118–198.
[96] Helffer, B., and Sjöstrand, J. Analyse semi-classique pour l’équation
de Harper III. Mémoire de la SMF No 39 ; Tome 117, Fasc. 4 (1989).
[97] Helffer, B., and Sjöstrand, J. Analyse semi-classique pour l’équation
de Harper II (comportement semi-classique près d’un rationnel). Mémoire de
la SMF No 40 ; Tome 118, Fasc. 1 (1990).
[98] Hofstadter, D. R. Energy levels and wave functions of Bloch electrons in
rational and irrational magnetic fields. Phys. Rev. B 14, 6 (1976), 2239–2249.
BIBLIOGRAPHIE 191
[99] Hou, J.-M. Light-induced Hofstadter’s butterfly spectrum of ultracold atoms

on the two-dimensional kagomé lattice. CHN. Phys. Lett. 26, 12 (2009).
[100] Ignat, R., and Millot, V. The critical velocity for vortex existence in
a two-dimensional rotating Bose-Einstein condensate. J. Funct. Anal. 233
(2006), 260–306.
[101] Ignat, R., and Millot, V. Energy expansion and vortex location for a two
dimensional rotating Bose-Einstein condensate. Rev. Math. Phys. 18 (2006),
119–162.
[102] Jaksch, D., and Zoller, P. Creation of effective magnetic fields in optical
lattices : the Hofstadter butterfly for cold neutral atoms. New J. Phys. 5, 56
(2003).
[103] Jerrard, R. L. Lower bounds for generalized Ginzburg-Landau functionals.
SIAM J. Appl. Math. 30 (1999), 721–746.
[104] Jerrard, R. L. Local minimizers with vortex filaments for a Gross-Pitaevsky
functional. ESAIM Contrôle Optim. Calc. Var. 13 (2007), 35–71.
[105] Jerrard, R. L., and Soner, H. M. The jacobian and the Ginzburg-Landau
energy. Calc. Var. Partial Differential Equations 14 (2002), 151–191.
[106] Jo, G.-B., Guzman, J., Thomas, C. K., Hosur, P., Vishwanath, A.,
and Stamper-Kurn, D. M. Ultracold atoms in a tunable optical kagome
lattice. Phys. Rev. Lett. 108, 4 (2012), 045305.
[107] Juzeliūnas, G., Ruseckas, J., Öhberg, P., and Fleischhauer, M.
Light-induced effective magnetic fields for ultracold atoms in planar geome-
tries. Phys. Rev. A 73, 025602 (2006).
[108] Karali, G. D., and Sourdis, C. The ground state of a Gross-Pitaevskii
energy with general potential in the Thomas-Fermi limit. Preprint on http:
//arxiv.org/abs/1205.5997, 2012.
[109] Kasamatsu, K., Tsubota, M., and Ueda, M. Vortices in multicomponent
Bose-Einstein condensates. Int. J. Mod. Phys. B 19, 1835 (2005).
[110] Kasamatsu, K., Yasui, Y., and Tsubota, M. Macroscopic quantum tun-
neling of two-component Bose-Einstein condensates. Phys. Rev. A 64, 053605
(2001).
[111] Kerdelhué, P. Équation de Schrödinger magnétique périodique avec symé-
trie d’ordre six. Mémoire de la SMF, 2ème série, tome 51, (1992), 1–139.
[112] Kerdelhué, P. Équation de Schrödinger magnétique périodique avec symé-
trie d’ordre six : mesure du spectre II. Annales de l’IHP (section Physique
théorique) 62, 2 (1995), 181–209.
[113] Ketterle, W. Nobel lecture : When atoms behave as waves : Bose-Einstein
condensation and the atom laser. Rev. Mod. Phys. 74, 4 (2002), 1131–1151.
[114] Lassoued, L., and Mironescu, P. Ginzburg-Landau type energy with
discontinuous constraint. J. Anal. Math. 77 (1999), 1–26.
[115] Lein, M. Weyl quantization and semiclassics. http://www-m5.ma.tum.de/
pers/Lein/lectures/quantisierung/quantization_semiclassics_2010.
06.30.pdf, 2010.
192 BIBLIOGRAPHIE
[116] Lieb, E. H. Thomas-Fermi and related theories of atoms and molecules. Rev.
Mod. Phys. 53, 4 (1981), 603–641.
[117] Lieb, E. H., and Loss, M. Analysis : Second Edition. No. 14 in Graduate
Studies in Mathematics. AMS, 2001.
[118] Lieb, E. H., Seiringer, R., Solovej, J., and Yngvason, J. The Mathe-
matics of the Bose Gas and its Condensation, vol. 34 of Oberwolfach Seminars.
Birkhäuser Basel, 2005.
[119] Lieb, E. H., Seiringer, R., and Yngvason, J. Bosons in a trap : A
rigorous derivation of the Gross-Pitaevskii energy functional. Phys. Rev. A
61, 043602 (2000).
[120] Lieb, E. H., Seiringer, R., and Yngvason, J. Derivation of the Gross-
Pitaevskii equation for rotating Bose gases. Comm. Math. Phys. 264, 2 (2006),
505–537.
[121] Lin, T.-C., and Wei, J. Ground state of N coupled nonlinear Schrödinger
equations in Rn , n ≤ 3. Comm. Math. Phys. 255, 3 (2005), 629–653.
[122] Lin, Y.-J., Compton, R. L., Jiménez-García, K., Porto, J. V., and
Spielman, I. B. Synthetic magnetic fields for ultracold neutral atoms. Nature
462 (2009), 628–632.
[123] Liu, Z. Phase separation of two-component Bose-Einstein condensates. J.
Math. Phys. 50, 10 (2009), 102104.
[124] Maddaloni, P., Modugno, M., Fort, C., Minardi, F., and Inguscio,
M. Collective oscillations of two colliding Bose-Einstein condensates. Phys.
Rev. Lett. 85, 12 (2000), 2413–2417.
[125] Maso, G. D. Integral representation on BV(ω) of Γ-limits of variational
integrals. Manuscripta Mathematica 30, 4 (1979), 387–416.
[126] Mason, P., and Aftalion, A. Classification of the ground states and topo-
logical defects in a rotating two-component Bose-Einstein condensate. Phys.
Rev. A 84, 3 (2011), 033611.
[127] Matthews, M. R., Hall, D. S., Jin, D. S., Ensher, J. R., Wieman,
C. E., and Cornell, E. A. Dynamical response of a Bose-Einstein conden-
sate to a discontinuous change in internal state. Phys. Rev. Lett. 81, 2 (1998),
243–247.
[128] McCarron, D. J., Cho, H. W., Jenkin, D. L., Köppinger, M. P., and
Cornish, S. L. Dual-species Bose-Einstein condensate of 87 Rb and 133 Cs.
Phys. Rev. A 84 (2011), 011603.
[129] Mekata, M. Kagome : The story of the basketweave lattice. Phys. Today Let-
ters. http://physicstoday.org/journals/doc/PHTOAD-ft/vol_56/iss_2/
12_1.shtml, February 2003.
[130] Modica, L. The gradient theory of phase transitions and the minimal inter-
face criterion. Arch. Rational Mech. Anal. 98, 2 (1987), 123–142.
[131] Modica, L., and Mortola, S. Un esempio di Γ-convergenza. Boll. Un.
Mat. Ital. B 5, 13 (1977), 285–299.
[132] Modugno, G., Modugno, M., Riboli, F., and G. Roati, a. M. I. Two
atomic species superfluid. Phys. Rev. Lett. 89, 19 (2002).
BIBLIOGRAPHIE 193
[133] Modugno, G., Modugno, M., Riboli, F., Roati, G., and Inguscio, M.
A two atomic species superfluid. Phys. Rev. Lett. 89 (2002), 190404–190408.
[134] Montambaux, G. Conduction quantique et physique mésoscopique. Pro-
gramme d’approfondissement Physique, avalaible in http://users.lps.
u-psud.fr/montambaux/polytechnique/PHY560B/PHY560B-2013.pdf ed.
École Polytechnique, 2013.
[135] Mumford, D., and Shah, J. Optimal approximations by piecewise smooth
functions and associated variational problems. Comm. Pure Appl. Math. 42,
5 (1989), 577–685.
[136] Myatt, C. J., Burt, E. A., Ghrist, R. W., Cornell, E. A., and Wie-
man, C. E. Production of two overlapping Bose-Einstein condensates by
sympathetic cooling. Phys. Rev. Lett. 78, 4 (1997), 586–589.
[137] Nobelprize.org. The Nobel Prize in Physics 1997 - Advanced Infor-
mation. http://www.nobelprize.org/nobel_prizes/physics/laureates/
1997/advanced.html.
2001/advanced.html.
2010/advanced.html.
[140] Noris, B., Tavares, H., Terracini, S., and Verzini, G. Uniform Hölder
bounds for nonlinear Schrödinger systems with strong competition. Comm.
Pure Appl. Math. 63, 3 (2010), 267–302.
[141] Novoselov, K. S. Nobel Lecture : Graphene : Materials in the Flatland.
Rev. Mod. Phys. 83, 3 (2011), 837–849.
[142] Öhberg, P., and Stenholm, S. Hartree-Fock treatment of the two-
component Bose-Einstein condensate. Phys. Rev. A 57, 2 (1998), 1272–1279.
[143] Panati, G., Spohn, H., and Teufel, S. Space-adiabatic perturbation
theory in quantum dynamics. Phys. Rev. Lett. 88, 250405 (2002).
[144] Panati, G., Spohn, H., and Teufel, S. Space-adiabatic perturbation
theory. Adv. Theor. Math. Phys. 7, 1 (2003), 145–204.
[145] Panati, G., Spohn, H., and Teufel, S. Motion of electrons in adiabatically
perturbed periodic structures. Analysis, Modeling and Simulation of Multiscale
Problems (2006), 595–617.
[146] Papp, S. B., Pino, J. M., and Wieman, C. E. Tunable miscibility in a
dual-species Bose-Einstein condensate. Phys. Rev. Lett. 101, 4 (2008), 040402.
[147] Persson, A. Bounds for the discrete part of the spectrum of a semi-bounded
Schrödinger operator. Math. Scand. 8 (1960), 143–153.
[148] Pethick, C. J., and Smith, H. Bose-Einstein Condensation in Dilute
Gases. Cambridge University Press, 2002.
[149] Phillips, W. D. Nobel Lecture : Laser cooling and trapping of neutral atoms.
Rev. Mod. Phys. 70, 3 (1998), 721–741.
194 BIBLIOGRAPHIE
[150] Pitaevskii, L., and Stringari, S. Bose Einstein Condensation, vol. 116.
International series of monographs on physics. Oxford Science Publications,
2003.
[151] Pólya, G., and Szegö, G. Isoperimetric inequalities in mathematical phy-
sics. Annals of Mathematics Studies, no. 27. Princeton University Press, 1951.
[152] Reed, M., and Simon, B. Methods of modern mathematical physics. IV.
Analysis of Operators. Academic Press, 1980.
[153] Royo-Letelier, J. Segregation and symmetry breaking of strongly coupled
two-component Bose-Einstein condensates in a harmonic trap. http://link.
springer.com/article/10.1007%2Fs00526-012-0571-7. To appear in Calc.
Var. Partial Differential Equations (2012).
[154] Ryu, C., Andersen, M. F., P. Cladé, V. N., Helmerson, K., and
Phillips, W. D. Observation of persistent flow of a Bose-Einstein condensate
in a toroidal trap. Phys. Rev. Lett. 99, 260401 (2007).
[155] Sandier, E. Lower bounds for the energy of unit vector fields and applica-
tions. J. Funct. Anal. 152 (Erratum : Ibid. 171 (2000), 233 1998), 379–403.
[156] Sandier, E., and Serfaty, S. Vortices in the Magnetic Ginzburg-Landau
Model, vol. 70 of Progress in Nonlinear Differential Equations and Their Ap-
plications. Birkhäuser, 2006.
[157] Santos, L., Baranov, M., Everts, J. C. J. H.-U., Fehrmann, H., and
Lewenstein, M. Atomic quatum gases in kagomé lattices. Phys. Rev. Lett.
93, 3 (2004).
[158] Sarvas, J. Symmetrization of condensers in n-space. Ann. Acad. Sci. Fenn.
Ser. A I, 522 (1972).
[159] Simon, B. Schrodinger operators in the twenty-first century. www.math.
caltech.edu/papers/bsimon/r40.ps.
[160] Simon, B. Semi-classical analysis of low lying eigenvalues I. Ann. Inst. H.
Poincaré 38 (1983), 295–307.
[161] Stock, S., Bretin, V., Chevy, F., and Dalibard, J. Shape oscillation
of a rotating Bose-Einstein condensate. Europhys. Lett. 65, 594 (2004).
[162] Syôzi, I. Statistics of kagomé lattice. Prog. Theor. Phys. 6, 3 (1951).
[163] Thalhammer, G., Barontini, G., Sarlo, L. D., Catani, J., Minardi,
F., and Inguscio, M. Double species Bose-Einstein condensate with tunable
interspecies interactions. Phys. Rev. Lett. 100, 21 (2008), 210402.
[164] Wei, J., and Weth, T. Nonradial symmetric bound states for a system
of coupled Schrödinger equations. Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat.
Natur. Rend. Lincei (9) Mat. Appl. 18, 3 (2007), 279–293.
[165] Wei, J., and Weth, T. Asymptotic behaviour of solutions of planar elliptic
systems with strong competition. Nonlinearity 21, 2 (2008), 305–317.
[166] Wei, J., and Weth, T. Radial solutions and phase separation in a system of
two coupled Schrödinger equations. Arch. Rational Mech. Anal. 190, 1 (2008),
83–106.
BIBLIOGRAPHIE 195
[167] Weiler, C. N., Neely, T. W., Scherer, D. R., Bradley, A. S., Davis,
M. J., and Anderson, B. P. Spontaneous vortices in the formation of
Bose-Einstein condensates. Nature 455, 7215 (2008), 948–951.
[168] Wilkinson, M. Critical properties of electron eigenstates in incommensurate
systems. Proc. Roy. Soc. Lond. A391 (1984), 305–350.
[169] Y.Xiao, Pelletier, V., Chaikin, P. M., and Huse, D. A. Landau levels
in the case of two degenerate coupled bands : kagome lattice tight-binding
spectrum. Phys. Rev. B 67, 10 (2003), 104505.
[170] Zak, J. Magnetic translation group. Physical Review 134, A1602-A1606
(1964).

Étude de Modèles Mathématiques Des Condensats de Bose-Einstein Pour Différents Types de Pièges Et D'interactions - Jimena Royo-Letelier

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Étude de Modèles Mathématiques Des Condensats de Bose-Einstein Pour Différents Types de Pièges Et D'interactions - Jimena Royo-Letelier

Transféré par

Droits d'auteur :

Formats disponibles

Étude de modèles mathématiques des

condensats de Bose-Einstein pour

THÈSE DE DOCTORAT DE L’UNIVERSITÉ DE

Pour obtenir le grade de docteur de l’Université de Versailles

Soutenue le 10 juin 2013, devant le jury composé de :

Amandine AFTALION Directrice de thèse

¡M«as vale ser punky !

Il faut être toujours ivre.

Je voudrais remercier mes directeurs, Amandine Aftalion et Bernard Helffer, d’avoir

Au cours de ma thèse j’ai eu l’opportunité de visiter Robert L. Jerrard à l’Univer-

Pendant la dernière année j’ai été lauréate de la bourse Bourses L’Oréal-France -

Ce mémoire marque la fin de ma thèse de doctorat, mais il représente aussi l’abou-

Je laisse pour la fin les plus importants des remerciements.

Les mots me manquent pour exprimer ma reconnaissance envers Émilie. Je ne pour-

Enfin, je voudrais remercier ma famille, dans les mots de ma soeur Manuela, un

2 Non existence of vortices 49

3 Segregation and symmetry breaking 67

4 A phase transition problem 91

5 The magnetic Schrödinger operator 131

1.1 Condensats unidimensionnels . . . . . . . . . . . . . . . . . . . . . . 22

5.1 Square, triangular, hexagonal and kagome lattices . . . . . . . . . . . 133

1.1 La physique des condensats de Bose-Einstein

Pendant plusieurs décennies, la prédiction d’Einstein demeure une prévision théori-

L’expérience se déroule de la manière suivante : les atomes sont d’abord refroidis

L’observation du condensat est faite grâce à un éclairage lumineux, dont l’absorption

1.1.2 L’approximation de Hartree et l’énergie de

où pi = −i~∇xi est l’opérateur d’impulsion pour la i-ième particule. On commence

Ψ(x1 , . . . , xN ) = u(x1 ) · · · u(xN ) u ∈ L2 (R3 ; C) . (1.1.3)

On trouve N contributions identiques provenant du terme d’énergie cinétique et du

où Eg est l’équation de Gross-Pitaesvkii définie dans (1.1.1). Le nombre de particules

La démarche heuristique entreprise ci-dessus pour déduire l’énergie de Gross-Pitaevskii

1.1.3 Configurations expérimentales

Lz = −i~(x1 ∂x2 − x2 ∂x1 ) .

Pour obtenir un condensat en rotation, la vitesse Ω doit être inférieure à la fréquence

Lorsque Ω = 0, les minimiseurs de la fonctionnelle d’énergie (1.1.8) ont une phase

Réduction de dimension et périodicité

L’utilisation de lasers ou de potentiels magnétiques permet de confiner les atomes

façon similaire, un potentiel périodique unidimensionnel est généré en superposant

Figure 1.1 – Répresentation schématique des condensats unidimensionnels (gris)

Condensats de Bose Einstein à deux composants

Des condensats avec deux types de particules sont obtenus expérimentalement en

où Eg est l’énergie de Gross-Pitaevskii définie dans (1.1.1). Le paramètre g12 cor-

d’interaction entre les particules du premier composant (respectivement du deuxième

Les paramètres g1 et g2 peuvent être variés en modifiant la quantité des particules

Les simulations numériques d’un condensat à deux composants mis en rotation

1.1.4 Magnétisme artificiel avec des condensats de

La simulation de phénomènes quantiques avec des systèmes physiques classiques

Figure 1.2 – Papillon de Hofstadter. Cette figure représente le spectre d’énergies

1.2 La fonctionnelle de Gross-Pitaevskii

et sous la contrainte de masse

L’injection de H dans L2 (R2 ; C) est compacte lorsque V est un potentiel localement

de minimiseurs de Eg dans H sous la contrainte (1.2.3). Dans ce mémoire de thèse

V (x) = ω12 x21 + ω22 x22 . (1.2.5)

− ∆u + V (x)u + g|u|2 u = λg u , (1.2.6)

L’espace de minimisation H contient des fonctions à valeurs complexes, mais on peut

On étudie naturellement l’énergie Eg quand g est un petit ou un grand paramètre,

1.3 Régime de faible interaction et l’opérateur li-

On s’attend à que l’état de base du condensat soit proche de la première fonction

kug − cϕkL2 (R2 ) ≤ Cg /2

avec |c| = 1. De même, l’énergie du condensat est proche de la première valeur

1.4 Le régime de forte interaction et l’approxima-

V (x)u(x) + g|u(x)|2 u(x) = λu(x) ,

et on effectue un changement d’échelle en posant