Variational Methods

Ergebnisse der Mathematik und ihrer Grenzgebiete 3. Folge A Series of Modern Surveys in Mathematics Editorial Board M...

Author: Michael Struwe

132 downloads 2351 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Ergebnisse der Mathematik und ihrer Grenzgebiete 3. Folge

A Series of Modern Surveys in Mathematics

Editorial Board M. Gromov, Bures-sur-Yvette J. Jost, Leipzig J. Kollár, Princeton G. Laumon, Orsay H. W. Lenstra, Jr., Leiden J. Tits, Paris D. B. Zagier, Bonn G. Ziegler, Berlin Managing Editor R. Remmert, Münster

Volume 34

Michael Struwe

Variational Methods Applications to Nonlinear Partial Differential Equations and Hamiltonian Systems Fourth Edition

123

Michael Struwe ETH Zürich Departement Mathematik Rämistr. 101 8092 Zürich, Switzerland

ISBN 978-3-540-74012-4

e-ISBN 978-3-540-74013-1

DOI 10.1007/978-3-540-74013-1 Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics ISSN 0071-1136 Library of Congress Control Number: 2008923744 Mathematics Subject Classification (2000): 58E05, 58E10, 58E12, 58E30, 58E35, 34C25, 34C35, 35A15, 35K15, 35K20, 35K22, 58F05, 58F22, 58G11 © 2008 Springer-Verlag Berlin Heidelberg This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: by the author using a Springer TEX macro package Production: LE-TEX Jelonek, Schmidt & Vöckler GbR, Leipzig Cover design: WMX Design GmbH, Heidelberg Printed on acid-free paper 9 8 7 6 5 4 3 2 1 springer.com

Preface to the Fourth Edition

Almost twenty years after conception of the ﬁrst edition, it was a challenge to prepare an updated version of this text on the Calculus of Variations. The ﬁeld has truely advanced dramatically since that time, to an extent that I ﬁnd it impossible to give a comprehensive account of all the many important developments that have occurred since the last edition appeared. Fortunately, an excellent overview of the most signiﬁcant results, with a focus on functional analytic and Morse theoretical aspects of the Calculus of Variations, can be found in the recent survey paper by Ekeland-Ghoussoub [1]. I therefore have only added new material directly related to the themes originally covered. Even with this restriction, a selection had to be made. In view of the fact that ﬂow methods are emerging as the natural tool for studying variational problems in the ﬁeld of Geometric Analysis, an emphasis was placed on advances in this domain. In particular, the present edition includes the proof for the convergence of the Yamabe ﬂow on an arbitrary closed manifold of dimension 3 ≤ m ≤ 5 for initial data allowing at most single-point blow-up. Moreover, we give a detailed treatment of the phenomenon of blow-up and discuss the newly discovered results for backward bubbling in the heat ﬂow for harmonic maps of surfaces. Aside from these more signiﬁcant additions, a number of smaller changes have been made throughout the text, thereby taking care not to spoil the freshness of the original presentation. References have been updated, whenever possible, and several mistakes that had survived the past revisions have now been eliminated. I would like to thank Silvia Cingolani, Irene Fonseca, Emmanuel Hebey, and Maximilian Schultz for helpful comments in this regard. Moreover, I am indebted to Gilles Angelsberg, Ruben Jakob, Reto M¨ uller, and Melanie Rupﬂin, for carefully proof-reading the new material. Z¨ urich, July 2007

Michael Struwe

Preface to the Third Edition

The Calculus of Variations continues to be an area of very rapid growth. Variational methods are indispensable as a tool in mathematical physics and geometry. Results on Ginzburg-Landau type variational problems inspire research on the related Seiberg-Witten functional on a K¨ ahler surface and invite speculations about possible applications in topology (Ding-Jost-Li-Peng-Wang [1]). Variational methods are applied in cosmology, as in the recent work of Fortunato-Giannoni-Masiello [1] and Giannoni-Masiello-Piccione [1] on geodesics in Lorentz manifolds and gravitational lenses. Applications to Hamiltonian dynamics now include a proof of the Seifert conjecture on brake orbits (Giannoni [1]) and results on homoclinic and heteroclinic solutions (Coti Zelati-Ekeland-S´er´e [1], Rabinowitz [1], S´er´e [1]) with interesting counterparts in the ﬁeld of semilinear elliptic equations (Coti-ZelatiRabinowitz [1], Rabinowitz [13]). The Calculus of Variations also has advanced on a more technical level. Campa-Degiovanni [1], Corvellec-Degiovanni-Marzocchi [1], Degiovanni-Marzocchi [1], Ioﬀe [1], and Ioﬀe-Schwartzman [1] have extended critical point theory to functionals on metric spaces, with applications, for instance, to quasilinear elliptic equations (Arioli [1], Arioli-Gazzola [1], Canino-Degiovanni [1]). Bolle [1] has proposed a new approach to perturbation theory, as treated in Section II.7 of this monograph. Numerous applications are studied in BolleGhoussoub-Tehrani [1]. The method of parameter dependence as in Sections I.7 and II.9 has found further striking applications in Chern-Simons theory (Struwe-Tarantello [1]) and independently for a related problem in mean ﬁeld theory (Ding-Jost-LiWang [1]). Inspired by these results, Wang-Wei [1] were able to solve a problem in chemotaxis with a similar structure. Jeanjean [1] and Jeanjean-Toland [1] have discovered an abstract setting where parameter dependence may be exploited. Ambrosetti [1], Ambrosetti-Badiale-Cingolani [1], and Ambrosetti-Badiale [1], [2] have found new applications of variational methods in bifurcation theory, reﬁning the classical results of B¨ohme [1] and Marino [1]. In Ambrosetti-Garcia Azorero-Peral [1] these ideas are applied to obtain precise existence results for conformal metrics of prescribed scalar curvature close to a constant, which shed new light on the work of Bahri-Coron [1], [2], Chang-Yang [1] quoted in Section III.4.11. The ﬁeld of critical equations as in Chapter III has been particularly active. Concentration proﬁles for Palais-Smale sequences as in Theorem III.3.1 have been studied in more detail by Rey [1] and Flucher [1].

viii

Preface to the Third Edition

Quite surprisingly, results analogous to Theorem III.3.1 have been discovered also for sequences of solutions to critical semilinear wave equations (Bahouri-G´erard [1]). For the semilinear elliptic equations of critical exponential growth related to the Moser-Trudinger inequality on a planar domain the patterns for existence and non-existence results are strikingly analogous to the higher dimensional case (Adimurthi [1], Adimurthi-Srikanth-Yadava [1]), and, on a macroscopic scale, quantization phenomena analogous to Theorem III.3.1 are observed for concentrating solutions of semilinear equations with exponential growth (Brezis-Merle [1], Li-Shafrir [1]). However, results of Struwe [17] and Ogawa-Suzuki [1] on the one hand and an example by Adimurthi-Prashanth [1] on the other suggest that there may be many qualitatively distinct types of blow-up behavior for Palais-Smale sequences in this case. Still, Theorem III.3.1 remains valid for solutions (Adimurthi-Struwe [1]) and also the analogue of Theorem III.3.4 has been obtained (Struwe [25]). The many similarities and subtle diﬀerences to the critical semilinear equations in higher dimensions make this ﬁeld particularly attractive for further study. References have been updated and a small number of mistakes have been rectiﬁed. I am indepted to Gerd M¨ uller, Paul Rabinowitz, and Henry Wente for their comments. Z¨ urich, July 1999

Michael Struwe

Preface to the Second Edition

During the short period of ﬁve years that have elapsed since the publication of the ﬁrst edition a number of interesting mathematical developments have taken place and important results have been obtained that relate to the theme of this book. First of all, as predicted in the Preface to the ﬁrst edition, Morse theory, indeed, has gone through a dramatic change, inﬂuenced by the work by Andreas Floer on Hamiltonian systems and in particular, on the Arnold conjecture. There are now also excellent accounts of these developments and their ramiﬁcations; see, in particular, the monograph by Matthias Schwarz [1]. The book by Hofer-Zehnder [2] on Symplectic Geometry shows that variational methods and, in particular, Floer theory have applications that range far beyond the classical area of analysis. Second, as a consequence of an observation by Stefan M¨ uller [1] which prompted the seminal work of Coifman-Lions-Meyer-Semmes [1], Hardy spaces and the space BMO are now playing a very important role in weak convergence results, in particular, when dealing with problems that exhibit a special (determinant) structure. A brief discussion of these results and some model applications can be found in Section I.3. Moreover, variational problems depending on some real parameter in certain cases have been shown to admit rather surprising a-priori bounds on critical points, with numerous applications. Some examples will be given in Sections I.7 and II.9. Other developments include the discovery of Hamiltonian systems with no periodic orbits on some given energy hypersurface, due to Ginzburg and Herman, and the discovery, by Chang-Ding-Ye, of ﬁnite time blow-up for the evolution problem for harmonic maps of surfaces, thus completing the results in Sections II.8, II.9 and III.6, respectively. A beautiful recent result of Ye concerns a new proof of the Yamabe theorem in the case of a locally conformally ﬂat manifold. This proof is presented in detail in Section III.4 of this new edition. In view of their numerous and wide-ranging applications, interest in variational methods is very strong and growing. Out of the large number of recent publications in the general ﬁeld of the calculus of variations and its applications some 50 new references have been added that directly relate to one of the themes in this monograph. Owing to the very favorable response with which the ﬁrst edition of this book was received by the mathematical community, the publisher has suggested that a second edition be published in the Ergebnisse series. It is a pleasure to thank all the many mathematicians, colleagues, and friends who

x

Preface to the Second Edition

have commented on the ﬁrst edition. Their enthusiasm has been highly inspiring. Moreover, I would like to thank, in particular, Matts Essen, Martin Flucher and Helmut Hofer for helpful suggestions in preparing this new edition. All additions and changes to the ﬁrst edition were carefully implemented by Suzanne Kronenberg, using the Springer TeX-Macros package, and I gratefully acknowledge her help. Z¨ urich, June 1996

Michael Struwe

Preface to the First Edition

It would be hopeless to attempt to give a complete account of the history of the calculus of variations. The interest of Greek philosophers in isoperimetric problems underscores the importance of “optimal form” already in ancient cultures; see Hildebrandt-Tromba [1] for a beautiful treatise of this subject. While variational problems thus are part of our classical cultural heritage, the ﬁrst modern treatment of a variational problem is attributed to Fermat, see Goldstine [1; p. 1]. Postulating that light follows a path of least possible time, in 1662 Fermat was able to derive the laws of refraction, thereby using methods which may already be termed analytic. With the development of the Calculus by Newton and Leibniz, the basis was laid for a more systematic development of the calculus of variations. The brothers Johann and Jakob Bernoulli and Johann’s student Leonhard Euler, all from the city of Basel in Switzerland, were to become the “founding fathers” (Hildebrandt-Tromba [1; p. 21]) of this new discipline. In 1743 Euler [1] submitted “A method for ﬁnding curves enjoying certain maximum or minimum properties”, published in 1744, the ﬁrst textbook on the calculus of variations. In an appendix to this book Euler [1; Appendix II, p. 298] expresses his belief that “every eﬀect in nature follows a maximum or minimum rule” (see also Goldstine [1; p. 106]), a credo in the universality of the calculus of variations as a tool. The same conviction also shines through Maupertuis’ [1] work on the famous “least action principle”, also published in 1744. (In retrospect, however, it seems that Euler was the ﬁrst to observe this important principle. See for instance Goldstine [1; p. 67 f. and p. 101 ﬀ.] for a more detailed historical account.) Euler’s book was a great source of inspiration for generations of mathematicians following. Major contributions were made by Lagrange, Legendre, Jacobi, Clebsch, Mayer, and Hamilton to whom we owe what we now call “Euler-Lagrange equations”, the “Jacobi diﬀerential equation” for a family of extremals, or “Hamilton-Jacobi theory”. The use of variational methods was not at all limited to one-dimensional problems in the mechanics of mass-points. In the 19th century variational methods also were employed for instance to determine the distribution of an electrical charge on the surface of a conductor from the requirement that the energy of the associated electrical ﬁeld be minimal (“Dirichlet’s principle”; see Dirichlet [1] or Gauss [1]) or were used in the construction of analytic functions (Riemann [1]). However, none of these applications was carried out with complete rigor. Often the model was confused with the phenomenon that it was supposed to describe and the fact (?) that for instance in nature there always exists an

xii

Preface to the First Edition

equilibrium distribution for an electrical charge on a conducting surface was taken as suﬃcient evidence for the corresponding mathematical problem to have a solution. A typical reasoning reads as follows: “In any event therefore the integral will be non-negative and hence there must exist a distribution (of charge) for which this integral assumes its minimum value,” (Gauss [1; p. 232], translation by the author). However, towards the end of the 19th century progress in abstraction and a better understanding of the foundations of the calculus opened such arguments to criticism. Soon enough, Weierstrass [1; pp. 52–54] found an example of a variational problem that did not admit a minimum solution. Weierstrass challenged his colleagues to ﬁnd a continuously diﬀerentiable function u: [−1, 1] → IR minimizing the integral 1 d 2 I(u) = x dx u dx −1

subject (for instance) to the boundary conditions u(±1) = ±1. Choosing uε (x) =

arctan( xε ) , ε > 0, arctan( 1ε )

as a family of comparison functions, Weierstrass was able to show that the inﬁnium of I in the above class was 0; however, the value 0 is not attained. (See also Goldstine [1; p. 371 f.].) Weierstrass’ critique of Dirichlet’s principle precipitated the calculus of variations into a Grundlagenkrise comparable to the crisis in set theory and logic after Russel’s discovery of antinomies in Cantor’s set theory or G¨ odel’s incompleteness proof. However, through the combined eﬀorts of several mathematicians who did not want to give up the wonderful tool that Dirichlet’s principle had been – including Weierstrass, Arz´ela, Fr´echet, Hilbert, and Lebesgue – the calculus of variations was revalidated and emerged from its crisis with new strength and vigor. Hilbert’s speech at the centennial assembly of the International Congress 1900 in Paris, where he proposed his famous 20 problems – two of which were devoted to questions related to the calculus of variatons – marks this newly found conﬁdence. In fact, following Hilbert’s [1] and Lebesgue’s [1] solution of the Dirichlet problem, a development began which within a few decades brought tremendous success, highlighted by the 1929 theorem of Ljusternik and Schnirelman [1] on the existence of three distinct prime closed geodesics on any compact surface of genus zero, or the 1930/31 solution of Plateau’s problem by Douglas [1], [2] and Rad` o [1]. The Ljusternik-Schnirelman result (and a previous result by Birkhoﬀ [1], proving the existence of one closed geodesic on a surface of genus 0) also marks the beginning of global analyis. This goes beyond Dirichlet’s principle as we no longer consider only minimizers (or maximizers) of variational

Preface to the First Edition

xiii

integrals, but instead look at all their critical points. The work of Ljusternik and Schnirelman revealed that much of the complexity of a function space is invariably reﬂected in the set of critical points of any variational integral deﬁned on it, an idea whose importance for the further development of mathematics can hardly be overestimated, whose implications even today may only be conjectured, and whose applications seem to be virtually unlimited. Later, Ljusternik and Schnirelman [2] laid down the foundations of their method in a general theory. In honor of their pioneering eﬀort any method which seeks to draw information concerning the number of critical points of a functional from topological data today often is referred to as Ljusternik-Schnirelman theory. Around the time of Ljusternik and Schnirelman’s work, another – equally important – approach towards a global theory of critical points was pursued by Marston Morse [2]. Morse’s work also reveals a deep relation between the topology of a space and the number and types of critical points of any function deﬁned on it. In particular, this led to the discovery of unstable minimal surfaces through the work of Morse-Tompkins [1], [2] and Shiﬀman [1], [2]. Somewhat reshaped and clariﬁed, in the 50’s Morse theory was highly successful in topology (see Milnor [1] and Smale [1]). After Palais [1], [2] and Smale [2] in the 60’s succeeded in generalizing Milnor’s constructions to inﬁnite-dimensional Hilbert manifolds – see also Rothe [1] for some early work in this regard – Morse theory ﬁnally was recognized as a useful (and usable) instrument also for dealing with partial diﬀerential equations. However, applications of Morse theory seemed somewhat limited in view of prohibitive regularity and non-degeneracy conditions to be met in a variational problem, conditions which – by the way – were absent in Morse’s original work. Today, inspired by the deep work of Conley [1], Morse theory seems to be turning back to its origins again. In fact, a Morse-Conley theory is emerging which one day may provide a tool as universal as Ljusternik-Schnirelman theory and still oﬀer an even better resolution of the relation between the critical set of a functional and topological properties of its domain. However, in spite of encouraging results, for instance by Benci [4], Conley-Zehnder [1], JostStruwe [1], Rybakowski [1], [2], Rybakowski-Zehnder [1], Salamon [1], and – in particular – Floer [1], a general theory of this kind does not yet exist. In these notes we want to give an overview of the state of the art in some areas of the calculus of variations. Chapter I deals with the classical direct methods and some of their recent extensions. In Chapters II and III we discuss minimax methods, that is, Ljusternik-Schnirelman theory, with an emphasis on some limiting cases in the last chapter, leaving aside the issue of Morse theory whose face is currently changing all too rapidly. Examples and applications are given to semilinear elliptic partial diﬀerential equations and systems, Hamiltonian systems, nonlinear wave equations, and problems related to harmonic maps of Riemannian manifolds or surfaces of prescribed mean curvature. Although our selection is of course biased by the interests of the author, an eﬀort has been made to achieve a good balance between diﬀerent areas of current research. Most of the results are known;

xiv

Preface to the First Edition

some of the proofs have been reworked and simpliﬁed. Attributions are made to the best of the author’s knowledge. No attempt has been made to give an exhaustive account of the ﬁeld or a complete survey of the literature. General references for related material are Berger-Berger [1], Berger [1], Chow-Hale [1], Eells [1], Nirenberg [1], Rabinowitz [11], Schwartz [2], Zeidler [1]; in particular, we recommend the recent books by Ekeland [2] and MawhinWillem [1] on variational methods with a focus on Hamiltonian systems and the forthcoming works of Chang [7] and Giaquinta-Hildebrandt. Besides, we mention the classical textbooks by Krasnoselskii [1] (see also KrasnoselskiiZabreiko [1]), Ljusternik-Schnirelman [2], Morse [2], and Vainberg [1]. As for applications to Hamiltonian systems and nonlinear variational problems, the interested reader may also ﬁnd additional references on a special topic in these ﬁelds in the short surveys by Ambrosetti [2], Rabinowitz [9], or Zehnder [1]. The material covered in these notes is designed for advanced graduate or Ph.D. students or anyone who wishes to acquaint himself with variational methods and possesses a working knowledge of linear functional analysis and linear partial diﬀerential equations. Being familiar with the deﬁnitions and basic properties of Sobolev spaces as provided for instance in the book by Gilbarg-Trudinger [1] is recommended. However, some of these prerequisites can also be found in the appendix. In preparing this manuscript I have received help and encouragement from a number of friends and colleagues. In particular, I wish to thank Proﬀ. Herbert Amann and Hans-Wilhelm Alt for helpful comments concerning the ﬁrst two sections of Chapter I. Likewise, I am indebted to Prof. J¨ urgen Moser for useful suggestions concerning Section I.4 and to Proﬀ. Helmut Hofer and Eduard Zehnder for advice on Sections I.6, II.5, and II.8, concerning Hamiltonian systems. Moreover, I am grateful to Gabi Hitz, Peter Bamert, Jochen Denzler, Martin Flucher, Frank Josellis, Thomas Kerler, Malte Sch¨ unemann, Miguel Sofer, Jean-Paul Theubet, and Thomas Wurms for going through a set of preliminary notes for this manuscript with me in a seminar at ETH Z¨ urich during the winter term of 1988/89. The present text certainly has proﬁted a great deal from their careful study and criticism. Special thanks I also owe to Kai Jenni for the wonderful typesetting of this manuscript with the TEX text processing system. I dedicate this book to my wife Anne. Z¨ urich, January 1990

Michael Struwe

Contents

Chapter I. The Direct Methods in the Calculus of Variations . . . . . . . . . .

1

1. Lower Semi-continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2

Degenerate Elliptic Equations, 4 — Minimal Partitioning Hypersurfaces, 6 — Minimal Hypersurfaces in Riemannian Manifolds, 7 — A General Lower Semi-continuity Result, 8

2. Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

Semilinear Elliptic Boundary Value Problems, 14 — Perron’s Method in a Variational Guise, 16 — The Classical Plateau Problem, 19

3. Compensated Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

Applications in Elasticity, 29 — Convergence Results for Nonlinear Elliptic Equations, 32 — Hardy Space Methods, 35

4. The Concentration-Compactness Principle . . . . . . . . . . . . . . . . . . . . . . .

36

Existence of Extremal Functions for Sobolev Embeddings, 42

5. Ekeland’s Variational Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

Existence of Minimizers for Quasi-convex Functionals, 54

6. Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

58

Hamiltonian Systems, 60 — Periodic Solutions of Nonlinear Wave Equations, 65

7. Minimization Problems Depending on Parameters . . . . . . . . . . . . . . .

69

Harmonic Maps with Singularities, 71

Chapter II. Minimax Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

74

1. The Finite Dimensional Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

74

2. The Palais-Smale Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

3. A General Deformation Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

Pseudo-gradient Flows on Banach Spaces, 81 — Pseudo-gradient Flows on Manifolds, 85

4. The Minimax Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Closed Geodesics on Spheres, 89

87

xvi

Contents

5. Index Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

94

Krasnoselskii Genus, 94 — Minimax Principles for Even Functionals, 96 — Applications to Semilinear Elliptic Problems, 98 — General Index Theories, 99 — Ljusternik-Schnirelman Category, 100 — A Geometrical S 1 -Index, 101 — Multiple Periodic Orbits of Hamiltonian Systems, 103

6. The Mountain Pass Lemma and its Variants . . . . . . . . . . . . . . . . . . . . .

108

Applications to Semilinear Elliptic Boundary Value Problems, 110 — The Symmetric Mountain Pass Lemma, 112 — Application to Semilinear Equations with Symmetry, 116

7. Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

118

Applications to Semilinear Elliptic Equations, 120

8. Linking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

125

Applications to Semilinear Elliptic Equations, 128 — Applications to Hamiltonian Systems, 130

9. Parameter Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

137

10. Critical Points of Mountain Pass Type . . . . . . . . . . . . . . . . . . . . . . . . . . .

143

Multiple Solutions of Coercive Elliptic Problems, 147

11. Non-diﬀerentiable Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

150

12. Ljusternik-Schnirelman Theory on Convex Sets . . . . . . . . . . . . . . . . . .

162

Applications to Semilinear Elliptic Boundary Value Problems, 166

Chapter III. Limit Cases of the Palais-Smale Condition . . . . . . . . . . . . . . .

169

1. Pohoˇzaev’s Non-existence Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

170

2. The Brezis-Nirenberg Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

173

Constrained Minimization, 174 — The Unconstrained Case: Local Compactness, 175 — Multiple Solutions, 180

3. The Eﬀect of Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

183

A Global Compactness Result, 184 — Positive Solutions on Annular-Shaped Regions, 190

4. The Yamabe Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Variational Approach, 195 — The Locally Conformally Flat Case, 197 — The Yamabe Flow, 198 — The Proof of Theorem 4.9 (following Ye [1]), 200 — Convergence of the Yamabe Flow in the General Case, 204 — The Compact Case u∞ > 0, 211 — Bubbling: The Case u∞ ≡ 0 , 216

194

Contents

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

xvii

220

Small Solutions, 221 — The Volume Functional, 223 — Wente’s Uniqueness Result, 225 — Local Compactness, 226 — Large Solutions, 229

6. Harmonic Maps of Riemannian Surfaces . . . . . . . . . . . . . . . . . . . . . . . . .

231

The Euler-Lagrange Equations for Harmonic Maps, 232 — Bochner identity, 234 — The Homotopy Problem and its Functional Analytic Setting, 234 — Existence and Non-existence Results, 237 — The Heat Flow for Harmonic Maps, 238 — The Global Existence Result, 239 — The Proof of Theorem 6.6, 242 — Finite-Time Blow-Up, 253 — Reverse Bubbling and Nonuniqueness, 257

Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

263

Sobolev Spaces, 263 — H¨ older Spaces, 264 — Imbedding Theorems, 264 — Density Theorem, 265 — Trace and Extension Theorems, 265 — Poincar´e Inequality, 266

Appendix B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

268

Schauder Estimates, 268 — Lp -Theory, 268 — Weak Solutions, 269 — A Regularity Result, 269 — Maximum Principle, 271 — Weak Maximum Principle, 272 — Application, 273

Appendix C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

274

Fr´ echet Diﬀerentiability, 274 — Natural Growth Conditions, 276

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

277

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

301

Glossary of Notations

V, V ∗ · · ∗ ·, ·: V × V ∗ → IR

generic Banach space with dual V ∗ norm in V induced norm in V ∗ , often also denoted · dual pairing, occasionally also used to denote scalar product in IRn E generic energy functional DE Fr´echet derivative Dom(E) domain of E v, DE(u) = DE(u)v = Dv E(u) directional derivative of E at u in direction v Lp (Ω; IRn ) space of Lebesgue-measurable functions u: Ω → IRn with ﬁnite Lp -norm 1/p uLp = |u|p dx , 1≤p<∞ Ω

L∞ (Ω; IRn )

space of Lebesgue-measurable and essentially bounded functions u: Ω → IRn with norm uL∞ = ess sup |u(x)| x∈Ω

H m,p (Ω; IRn )

H0m,p (Ω; IRn )

H −m,q (Ω; IRn ) Dm,p (Ω; IRn )

Sobolev space of functions u ∈ Lp (Ω; IRn ) with |∇k u| ∈ Lp (Ω) for all k ∈ INn0 , |k| ≤ m, with norm uH m,p = 0≤|k|≤m ∇k uLp

completion of C0∞ (Ω; IRn ) in the norm · H m,p ; if Ω is bounded an equivalent norm is given by uH0m,p = |k|=m ∇k uLp

dual of H0m,p (Ω; IRn ), where p1 = 1q = 1; q is omitted, if p = q = 2 completion of C0∞ (Ω; IRn ) in the norm uDm,p = k p |k|=m ∇ uL

xx

Glossary of Notations

C m,α (Ω; IRn )

space of m times continuously diﬀerentiable functions u: Ω → IRn whose mth order derivatives are H¨ older continuous with exponent 0 ≤ α ≤ 1 space of smooth functions u: Ω → IRn with compact C0∞ (Ω; IRn ) support in Ω supp(u) = {x ∈ Ω ; u(x) = 0} support of a function u: Ω → IRn Ω ⊂⊂ Ω the closure of Ω is compact and contained in Ω restriction of a measure Ln Lebesgue measure on IRn Bρ (u; V ) = {v ∈ V ; u − v < ρ} open ball of radius ρ around u ∈ V ; in particular, if V = IRn , then Bρ (x0 ) = Bρ (x0 ; IRn ), Bρ = Bρ (0) Re real part Im imaginary part c, C generic constants Cross-references (N.x.y) refers to formula (x, y) in Chapter N (x.y) within Chapter N refers to formula (N.x.y)

Chapter I

The Direct Methods in the Calculus of Variations

Many problems in analysis can be cast into the form of functional equations F (u) = 0, the solution u being sought among a class of admissible functions belonging to some Banach space V . Typically, these equations are nonlinear; for instance, if the class of admissible functions is restricted by some (nonlinear) constraint. A particular class of functional equations is the class of Euler-Lagrange equations DE(u) = 0 for a functional E on V , which is Fr´echet-diﬀerentiable with derivative DE. We say such equations are of variational form. For equations of variational form an extensive theory has been developed, and variational principles play an important role in mathematical physics and diﬀerential geometry, optimal control and numerical anlysis. We brieﬂy recall the basic deﬁnitions that will be needed in this and the following chapters, see Appendix C for details: Suppose E is a Fr´echet-diﬀerentiable functional on a Banach space V with normed dual V ∗ and duality pairing ·, · : V × V ∗ → IR, and let DE : V → V ∗ denote the Fr´echet-derivative of E. Then the directional (Gateaux-) derivative of E at u in the direction of v is given by d E(u + εv) = v, DE(u) = DE(u) v. dε ε=0 For such E, we call a point u ∈ V critical if DE(u) = 0; otherwise, u is called regular. A number β ∈ IR is a critical value of E if there exists a critical point u of E with E(u) = β. Otherwise, β is called regular. Of particular interest (also in the non-diﬀerentiable case) will be relative minima of E, possibly subject to constraints. Recall that for a set M ⊂ V a point u ∈ M is an absolute minimizer for E on M if for all v ∈ M there holds E(v) ≥ E(u). A point u ∈ M is a relative minimizer for E on M if for some neighborhood U of u in V it is absolutely E-minimizing in M ∩ U . Moreover, in the diﬀerentiable case, we shall also be interested in the existence of saddle points, that is, critical points u of E such that any neighborhood U of u in V contains points v, w such that E(v) < E(u) < E(w). In physical systems, saddle points appear as unstable equilibria or transient excited states.

2

Chapter I. The Direct Methods in the Calculus of Variations

In this chapter we review some basic methods for proving the existence of relative minimizers. Somewhat imprecisely we summarily refer to these methods as the direct methods in the calculus of variations. However, besides the classical lower semi-continuity and compactness method we also include the compensated compactness method of Murat and Tartar, and the concentrationcompactness principle of P.L. Lions. Moreover, we recall Ekeland’s variational principle and the duality method of Clarke and Ekeland. Applications will be given to problems concerning minimal hypersurfaces, semilinear and quasi-linear elliptic boundary value problems, ﬁnite elasticity, Hamiltonian systems, and semilinear wave equations. From the beginning it will be apparent that in order to achieve a satisfactory existence theory the notion of solution will have to be suitably relaxed. Hence, in general, the above methods will at ﬁrst only yield generalized or “weak” solutions of our problems. A second step often will be necessary to show that these solutions are regular enough to be admitted as classical solutions. The regularity theory in many cases is very subtle and involves a delicate machinery. It would go beyond the scope of this book to cover this topic completely. However, for the problems that we will mostly be interested in, the regularity question can be dealt with rather easily. The reader will ﬁnd this material in Appendix B. References to more advanced texts on the regularity issue will be given where appropriate.

1. Lower Semi-continuity In this section we give suﬃcient conditions for a functional to be bounded from below and to attain its inﬁmum. The discussion can be made largely independent of any diﬀerentiability assumptions on E or structure assumptions on the underlying space of admissible functions M . In fact, we have the following classical result. 1.1 Theorem. Let M be a topological Hausdorﬀ space, and suppose E : M → IR ∪ +∞ satisﬁes the condition of bounded compactness:

(1.1)

For any α ∈ IR the set Kα = {u ∈ M ; E(u) ≤ α} is compact (Heine-Borel property).

Then E is uniformly bounded from below on M and attains its inﬁmum. The conclusion remains valid if instead of (1.1) we suppose that any sub-level set Kα is sequentially compact.

1. Lower Semi-continuity

3

Remark. Necessity of condition (1.1) is illustrated by simple examples: The function E: [−1, 1] → IR given by E(x) = x2 if x = 0, E(x) = 1 if x = 0, or the exponential function E(x) = exp(x) on IR are bounded from below but do not admit a minimizer. Note that the space M in the ﬁrst example is compact while in the second example the function E is smooth – even analytic. Proof of Theorem 1.1. Suppose (1.1) holds. We may assume E ≡ +∞. Let α0 = inf E ≥ −∞, M

and let (αm ) be the strictly decreasing sequence αm α0

(m → ∞) .

Let Km = Kαm . By assumption, each Km is compact and non-empty. Moreover,Km ⊃ Km+1 for all m. By compactness of Km there exists a point u ∈ m∈IN Km , satisfying E(u) ≤ αm ,

for all m.

Passing to the limit m → ∞ we obtain that E(u) ≤ α0 = inf E, M

and the claim follows. If instead of (1.1) each Kα is sequentially compact, we choose a minimizing sequence (um ) in M such that E(um ) → α0 . Then for any α > α0 the sequence (um ) will eventually lie entirely within Kα . Bysequential compactness of Kα therefore (um ) will accumulate at a point u ∈ α>α0 Kα which is the desired minimizer. Note that if E : M → IR satisﬁes (1.1), then for any α ∈ IR the set {u ∈ M ; E(u) > α} = M \ Kα is open, that is, E is lower semi-continuous. (Respectively, if each Kα is sequentially compact, then E will be sequentially lower semi-continuous.) Conversely, if E is (sequentially) lower semi-continuous and for some α ∈ IR the set Kα is (sequentially) compact, then Kα will be (sequentially) compact for all α ≤ α and again the conclusion of Theorem 1.1 will be valid. Note that the lower semi-continuity condition can be more easily fulﬁlled the ﬁner the topology on M . In contrast, the condition of compactness of the sub-level sets Kα , α ∈ IR, calls for a coarse topology and both conditions are competing. In practice, there is often a natural weak Sobolev space topology where both conditions can be simultaneously satisﬁed. However, there are many interesting cases where condition (1.1) cannot hold in any reasonable topology (even though relative minimizers may exist). Later in this chapter we

4

Chapter I. The Direct Methods in the Calculus of Variations

shall see some examples and some more delicate ways of handling the possible loss of compactness. See Section 4; see also Chapter III. In applications, the conditions of the following special case of Theorem 1.1 can often be checked more easily. 1.2 Theorem. Suppose V is a reﬂexive Banach space with norm · , and let M ⊂ V be a weakly closed subset of V . Suppose E : M → IR ∪ +∞ is coercive and (sequentially) weakly lower semi-continuous on M with respect to V , that is, suppose the following conditions are fullﬁlled: (1◦) E(u) → ∞ as u → ∞, u ∈ M . (2◦) For any u ∈ M , any sequence (um ) in M such that um u weakly in V there holds: E(u) ≤ lim inf E(um ) . m→∞

Then E is bounded from below on M and attains its inﬁmum in M . The concept of minimizing sequences oﬀers a direct and (apparently) constructive proof. Proof. Let α0 = inf M E and let (um ) be a minimizing sequence in M , that is, satisfying E(um ) → α0 . By coerciveness, (um ) is bounded in V . Since V is ˇ reﬂexive, by the Eberlein-Smulian theorem (see Dunford-Schwartz [1; p. 430]) we may assume that um u weakly for some u ∈ V . But M is weakly closed, therefore u ∈ M , and by weak lower semi-continuity E(u) ≤ lim inf E(um ) = α0 . m→∞

Examples. An important example of a sequentially weakly lower semicontinous functional is the norm in a Banach space V . Closed and convex subsets of Banach spaces are important examples of weakly closed sets. If V is the dual of a separable normed vector space, Theorem 1.2 and its proof remain valid if we replace weak by weak∗ -convergence. We present some simple applications. Degenerate Elliptic Equations 1.3 Theorem. Let Ω be a bounded domain in IRn , p ∈ [2, ∞[ with conjugate exponent q satisfying p1 + 1q = 1, and let f ∈ H −1,q (Ω), the dual of H01,p (Ω), be given. Then there exists a weak solution u ∈ H01,p (Ω) of the boundary value problem (1.2) (1.3)

−∇ · (|∇u|p−2 ∇u) = f u=0

in the sense that u satisﬁes the equation

in Ω on ∂Ω

1. Lower Semi-continuity

(∇u|∇u|p−2 ∇ϕ − f ϕ )dx = 0 ,

(1.4) Ω

5

∀ϕ ∈ C0∞ (Ω) .

Proof. Note that the left part of (1.4) is the directional derivative of the C 1 functional 1 p |∇u| dx − f u dx E(u) = p Ω Ω on the Banach space V = H01,p (Ω) in direction ϕ; that is, problem (1.2), (1.3) is of variational form. Note that H01,p (Ω) is reﬂexive. Moreover, E is coercive. In fact, we have 1 1 upH 1,p − f H −1,q uH 1,p ≥ upH 1,p − cuH 1,p 0 0 p p 0 0 p −1 ≥ c uH 1,p − C.

E(u) ≥

0

Finally, E is (sequentially) weakly lower semi-continuous: It suﬃces to show that for um u weakly in H01,p (Ω) we have

f um dx → Ω

f u dx . Ω

Since f ∈ H −1,q (Ω) , however, this follows from the very deﬁnition of weak convergence. Hence Theorem 1.2 is applicable and there exists a minimizer u ∈ H01,p (Ω) of E, solving (1.4). Note that for p ≥ 2 the p-Laplacian is strongly monotone in the sense that Ω

|∇u|p−2 ∇u − |∇v|p−2 ∇v · (∇u − ∇v) dx ≥ cu − vpH 1,p . 0

In particular, the solution u to (1.4) is unique. If f is more regular, say f ∈ C m,α (Ω), we would expect the solution u of (1.4) to be more regular as well. This is true if p = 2, see Appendix B, but in the degenerate case p > 2, where the uniform ellipticity of the p-Laplace operator is lost at zeros of |∇u|, the best that one can hope for is u ∈ C 1,α (Ω); see Uhlenbeck [1], Tolksdorf [2; p. 128], Di Benedetto [1]. In Theorem 1.3 we have applied Theorem 1.2 to a functional on a reﬂexive space. An example in a non-reﬂexive setting is given next.

6

Chapter I. The Direct Methods in the Calculus of Variations

Minimal Partitioning Hypersurfaces For a domain Ω ⊂ IRn let BV (Ω) be the space of functions u ∈ L1 (Ω) such that

n |Du| = sup uDi gi dx ; Ω

Ω i=1

g = (g1 , . . . , gn ) ∈ C01 (Ω; IRn ), |g| ≤ 1 endowed with the norm

<∞,

uBV = uL1 +

|Du| . Ω

BV (Ω) is a Banach space, embedded in L1 (Ω), and – provided Ω is bounded and suﬃciently smooth – by Rellich’s theorem the injection BV (Ω) → L1 (Ω) is compact; see

for instance Giusti [1; Theorem 1.19, p. 17]. Moreover, the function u → Ω |Du| is lower semi-continuous with respect to L1 -convergence. Let χG be the characteristic function of a set G ⊂ IRn ; that is, χG (x) = 1 if x ∈ G, χG (x) = 0 else. Also let Ln denote the n-dimensional Lebesgue measure. 1.4 Theorem. Let Ω be a smooth, bounded domain in IRn . Then there exists a subset G ⊂ Ω such that (1◦ )

Ln (G) = Ln (Ω \ G) =

1 n L (Ω) 2

and such that its perimeter with respect to Ω, P (G, Ω) = |DχG | , (2◦ ) Ω

is minimal among all sets satisfying (1◦ ). Proof. Let M = {χG ; G ⊂ Ω is measurable and satisﬁes (1◦ )}, endowed with the L1 -topology, and let E : M → IR ∪ +∞ be given by E(u) = |Du| . Ω

Since χG L1 ≤ L (Ω), the functional E is coercive on M with respect to the norm in BV (Ω). Since bounded sets in BV (Ω) are relatively compact in L1 (Ω) and since M is closed in L1 (Ω), by weak lower semi-continuity of E in L1 (Ω) the sub-level sets of E are compact. The conclusion now follows from Theorem 1.1. n

The support of the distribution DχG , where G has minimal perimeter (2◦ ) with respect to Ω, can be interpreted as a minimal bisecting hypersurface,

1. Lower Semi-continuity

7

dividing Ω into two regions of equal volume. The regularity of the dividing hypersurface is intimately connected with the existence of minimal cones in IRn . See Giusti [1] for further material on functions of bounded variation, sets of bounded perimeter, the area integrand, and applications. A related setting for the study of minimal hypersurfaces and related objects is oﬀered by geometric measure theory. Also in this ﬁeld variational principles play an important role; see for instance Almgren [1], Morgan [1], or Simon [1] for introductory material and further references. Our next example is concerned with a parametric approach.

Minimal Hypersurfaces in Riemannian Manifolds Let Ω be a bounded domain in IRn , and let S be a compact subset in IRN . Also let u0 ∈ H 1,2 (Ω; IRN ) with u0 (Ω) ⊂ S be given. Deﬁne H 1,2 (Ω; S) = u ∈ H 1,2 (Ω; IRN ) ; u(Ω) ⊂ S almost everywhere and let

M = u ∈ H 1,2 (Ω; S) ; u − u0 ∈ H01,2 (Ω; IRN ) .

Then, by Rellich’s theorem, M is closed in the weak topology of V = H 1,2 (Ω; IRN ). For u = (u1 , . . . , uN ) ∈ H 1,2 (Ω; S) let gij (u)∇ui ∇uj dx ,

E(u) = Ω

where g = (gij )1≤i,j≤N is a given positively deﬁnite symmetric matrix with coeﬃcients gij (u) depending continuously on u ∈ S, and where, by convention, we tacitly sum over repeated indices 1 ≤ i, j ≤ N . Note that since S is compact g is uniformly positive deﬁnite on S, and there exists λ > 0 such that E(u) ≥ λ ∇u2L2 for u ∈ H 1,2 (Ω; S). In addition, since S and Ω are bounded, we have that uL2 ≤ c uniformly, for u ∈ H 1,2 (Ω; S). Hence E is coercive on H 1,2 (Ω; S) with respect to the norm in H 1,2 (Ω; IRN ). Finally, E is lower semi-continuous in H 1,2 (Ω; S) with respect to weak convergence in H 1,2 (Ω; IRN ). Indeed, if um u weakly in H 1,2 (Ω; IRN ), by Rellich’s theorem um → u strongly in L2 and hence a subsequence (um ) converges almost everywhere. By Egorov’s theorem, given δ > 0 there is an exceptional set Ωδ of measure Ln (Ωδ ) < δ such that um → u uniformly on Ω \ Ωδ . We may assume that Ωδ ⊂ Ωδ for δ ≤ δ . By weak lower semi-continuity of the semi-norm on H 1,2 (Ω; IRN ), deﬁned by |v|2 =

gij (u) ∇v i ∇v j dx, Ω\Ωδ

then

8

Chapter I. The Direct Methods in the Calculus of Variations

Ω\Ωδ

gij (u) ∇ui ∇uj dx gij (u) ∇uim ∇ujm dx ≤ lim inf m→∞ Ω\Ω δ gij (um ) ∇uim ∇ujm dx = lim inf m→∞

Ω\Ωδ

≤ lim inf E(um ) . m→∞

Passing to the limit δ → 0, from Beppo Levi’s theorem we obtain gij (u) ∇ui ∇uj dx E(u) = lim δ→0

Ω\Ωδ

≤ lim inf E(um ) . m→∞

Applying Theorem 1.2 to E on M we obtain 1.5 Theorem. For any boundary data u0 ∈ H 1,2 (Ω; S) there exists an Eminimal extension u ∈ M . In diﬀerential geometry Theorem 1.5 arises in the study of harmonic maps u : Ω → S from a domain Ω into an N -dimensional manifold S with metric g for prescribed boundary data u = u0 on ∂Ω. Like in the previous example, the regularity question is related to the existence of special harmonic maps; in this case, singularities of harmonic maps from Ω into S are related to harmonic mappings of spheres into S. For further references see Eells-Lemaire [1], [2], Hildebrandt [3], Jost [2]. For questions concerning regularity see GiaquintaGiusti [1], Schoen-Uhlenbeck [1], [2]. A General Lower Semi-continuity Result We now conclude this short list of introductory examples and return to the development of the variational theory. Note that the property of E being lower semi-continuous with respect to some weak kind of convergence is at the core of the above existence results. In Theorem 1.6 below we establish a lower semicontinuity result for a very broad class of variational integrals, including and going beyond those encountered in Theorem 1.5, as Theorem 1.6 would also apply in the case of unbounded targets S and possibly degenerate or singular metrics g. We consider variational integrals F (x, u, ∇u) dx (1.5) E(u) = Ω

involving (vector-valued) functions u : Ω ⊂ IRn → IRN .

1. Lower Semi-continuity

9

1.6 Theorem. Let Ω be a domain in IRn , and assume that F : Ω × IRN × IRnN → IR is a Caratheodory function satisfying the conditions (1◦) F (x, u, p) ≥ φ(x) for almost every x, u, p, where φ ∈ L1 (Ω). (2◦) F (x, u, ·) is convex in p for almost every x, u. 1,1 (Ω) and um → u in L1 (Ω ), ∇um ∇u weakly in Then, if um , u ∈ Hloc 1 L (Ω ) for all bounded Ω ⊂⊂ Ω, it follows that E(u) ≤ lim inf E(um ) , m→∞

where E is given by (1.5). Notes. In the scalar case N = 1, weak lower semi-continuity results like Theorem 1.6 were ﬁrst stated by Tonelli [1] and Morrey [1]; these results were then extended and simpliﬁed by Serrin [1], [2] who showed that for non-negative, smooth functions F (x, u, p): Ω × IR × IRn → IR, which are convex in p , the functional E given by (1.5) is lower semi-continuous with respect to convergence in L1loc (Ω). A corresponding result in the vector-valued case N > 1 subsequently was derived by Morrey [4; Theorem 4.1.1]; however, Eisen [1] not only pointed out a gap in Morrey’s proof but also gave an example showing that for N > 1 in general, Theorem 1.6 ceases to be true without the assumption that the L1 -norms of ∇um are uniformly locally bounded. Theorem 1.6 is due to Berkowitz [1] and Eisen [2]. Related results can be found for instance in Morrey [4; Theorem 1.8.2], or Giaquinta [1]. Our proof is modeled on Eisen [2]. Proof. We may assume that E(um ) is ﬁnite and convergent. Moreover, replacing F by F − φ we may assume that F ≥ 0. Let Ω ⊂⊂ Ω be given. By weak local L1 -convergence ∇um ∇u, for any m0 ∈ IN there exists a sequence (P l )l≥m0 of convex linear combinations Pl =

l

αlm ∇um , 0 ≤ αlm ≤ 1 ,

m=m0

l

αlm = 1 , l ≥ m0

m=m0

such that P l → ∇u strongly in L1 (Ω ) and pointwise almost everywhere as l → ∞; see for instance Rudin [1; Theorem 3.13]. By convexity, for any m0 , any l ≥ m0 , and almost every x ∈ Ω : l l l αm ∇um (x) F x, u(x), P (x) = F x, u(x), m=m0

≤

l

αlm F (x, u(x), ∇um(x)) .

m=m0

Integrating over Ω and passing to the limit l → ∞, from Fatou’s lemma we obtain:

10

Chapter I. The Direct Methods in the Calculus of Variations

F (x, u(x), ∇u(x)) dx ≤ lim inf F x, u(x), P l (x) dx l→∞ Ω Ω F x, u(x), ∇um (x) dx . ≤ sup m≥m0

Ω

Since m0 was arbitrary, this implies that Ω

F x, u(x), ∇u(x) dx ≤ lim sup m→∞

Ω

F (x, u(x), ∇um (x)) dx ,

for any bounded Ω ⊂⊂ Ω. Now we need the following result (Eisen [2; p. 75]). 1.7 Lemma. Under the hypotheses of Theorem 1.6 on F, um , and u there exists a subsequence (um ) such that: F (x, um (x), ∇um (x)) − F (x, u(x), ∇um(x)) → 0 in measure, locally in Ω. Proof of Theorem 1.6 (completed). By Lemma 1.7 for any Ω ⊂⊂ Ω,any ε> 0, <ε ⊂ Ω with Ln Ωε,m and any m0 ∈ IN there exists m ≥ m0 and a set Ωε,m such that (1.6)

|F x, um (x), ∇um (x) − F x, u(x), ∇um(x) | < ε

for all x ∈ Ω \ Ωε,m . Replacing ε by εm = 2−m and passing to a subsequence, if necessary, we may assume that for each m there is a set Ωε m ,m ⊂ Ω of measure < εm such that (1.6) is satisﬁed (with εm ) for all x ∈ Ω \ Ωε m,m . Hence, for any given ε > 0, if we choose m0 = m0 (ε) > | log2 ε|, Ωε = m≥m0 Ωε m ,m , this set has measure Ln (Ωε ) < ε and inequality (1.6) holds uniformly for all x ∈ Ω \ Ωε , and all m ≥ m0 (ε). Moreover, for ε < δ by construction Ωε ⊂ Ωδ . Cover Ω by disjoint bounded sets Ω (k) ⊂⊂ Ω, k ∈ IN. Let ε > 0 be given and choose a sequence ε(k) > 0, such that k∈IN Ln Ω (k) ε(k) ≤ ε. Passing (k) (k) (k) to a subsequence, if necessary, for each Ω and ε we may choose m0 and (k)

Ωε

(k)

⊂ Ω (k) such that Ln Ωε

< ε(k) and

|F (x, um (x), ∇um (x)) − F (x, u(x), ∇um (x)) | < ε(k) (k)

(k)

(k)

uniformly for x ∈ Ω (k) \Ωε , m ≥ m0 . Moreover, we may assume that Ωε ⊂ (k) (k) , ΩεK = Ωδ , if ε < δ, for all k. Then for any K ∈ IN, letting Ω K = ∪K k=1 Ω (k) K ∪k=1 Ωε , we have

1. Lower Semi-continuity

11

Ω K \ΩεK

F (x, u, ∇u) dx ≤ lim sup m→∞

Ω K \ΩεK

F (x, u, ∇um ) dx

≤ lim sup m→∞

Ω K \ΩεK

F (x, um , ∇um ) dx + ε

≤ lim sup E(um ) + ε = lim inf E(um ) + ε . m→∞

m→∞

Letting ε → 0 and then K → ∞, the claim follows from Beppo Levi’s theorem, since F ≥ 0 and since Ω K \ ΩεK increases when ε ↓ 0, followed by K ↑ ∞. Proof of Lemma 1.7. We basically follow Eisen [2]. Suppose by contradiction that there exist Ω ⊂⊂ Ω and ε > 0 such that, letting Ωm = {x ∈ Ω ; |F (x, um , ∇um ) − F (x, u, ∇um ) | ≥ ε} , there holds lim inf Ln (Ωm ) ≥ 2ε . m→∞

The sequence (∇um ), being weakly convergent, is uniformly bounded in L1 (Ω ). In particular, C ≤ε, Ln {x ∈ Ω ; |∇um (x)| ≥ l} ≤ l−1 |∇um | dx ≤ l Ω ˜ m := {x ∈ Ωm ; |∇um (x)| ≤ l0 (ε)} if l ≥ l0 (ε) is large enough. Setting Ω therefore there holds ˜m ≥ ε. lim inf Ln Ω Hence also for Ω M =

m→∞

˜m we have Ω

m≥M

Ln (Ω M ) ≥ ε , uniformlyin M ∈ IN. Moreover, Ω ⊃ Ω M ⊃ Ω M +1 for all M and therefore Ω ∞ := Ω M ⊂ Ω has Ln (Ω ∞ ) ≥ ε. Finally, neglecting a set of measure M ∈IN

zero and passing to a subsequence, if necessary, we may assume that F (x, z, p) is continuous in (z, p), that um (x), u(x), ∇um (x) are unambiguously deﬁned and ﬁnite while um (x) → u(x) as m → ∞ at every point x ∈ Ω ∞ . Note that every point x ∈ Ω ∞ by construction belongs to inﬁnitely many of ˜ ˜ m . Choose such a point x. Relabeling, we may assume x ∈ the sets Ω m∈IN Ωm .

12

Chapter I. The Direct Methods in the Calculus of Variations

By uniform boundedness |∇um (x)| ≤ C there exists a subsequence m → ∞ and a vector p ∈ IRnN such that ∇um (x) → p (m → ∞). But then by continuity F x, um (x), ∇um (x) → F x, u(x), p while also

F x, u(x), ∇um (x) → F x, u(x), p

which contradicts the characterization of Ωm given above. 1.8 Remarks. The following observations may be useful in applications. (1◦) Theorem 1.6 also applies to functionals involving higher (mth-) order derivatives of a function u by letting U = (u, ∇u, . . . , ∇m−1 u) denote the (m − 1)-jet of u. Note that convexity is only required in the highest-order derivatives P = ∇m u. (2◦) If (um ) is bounded in H 1,1 (Ω ) for any Ω ⊂⊂ Ω, by Rellich’s theorem and repeated selection of subsequences there exists a subsequence (um ) which converges strongly in L1 (Ω ) for any Ω ⊂⊂ Ω. Local boundedness in H 1,1 of a minimizing sequence (um ) for E can be inferred from a coerciveness condition like F (x, z, p) ≥ |p|μ − φ(x), μ ≥ 1, φ ∈ L1 .

(1.7)

The delicate part in the hypotheses concerning (um ) is the assumption that (∇um ) converges weakly in L1loc . In case μ > 1 in (1.7) this is clear, but in case μ = 1 the local L1 -limit of a minimizing sequence may lie in BVloc instead of 1,1 . See Theorem 1.4, for example; see also Section 3. Hloc ◦ (3 ) By convexity in p, continuity of F in (u, p) for almost every x is equivalent to the following condition, which is easier to check in applications: F (x, ·, ·) is continuous, separately in u ∈ IRN and p ∈ IRnN , for almost every x ∈ Ω. Indeed, for any ﬁxed x, u, p and all e ∈ IRnN , |e| = 1, α ∈ [0, 1], letting q = p + αe, p+ = p + e, p− = p − e and writing F (x, u, p) = F (u, p) for brevity, by convexity we have F (u, q) = F (u, αp+ + (1 − α)p) ≤ αF (u, p+ ) + (1 − α)F (u, p) , 1 α α 1 F (u, p) = F (u, q+ p− ) ≤ F (u, q) + F (u, p− ) . 1+α 1+α 1+α 1+α Hence α (F (u, p) − F (u, p+ )) ≤ F (u, p) − F (u, q) ≤ α (F (u, p− ) − F (u, p)) and it follows that |F (u, q) − F (u, p)| ≤ sup |F (u, q) − F (u, p)| . |q − p| |q−p|≤1 |q−p|=1 sup

2. Constraints

13

Since the sphere of radius 1 around p lies in the convex hull of ﬁnitely many vectors q0 , q1 , . . . , qnN , by continuity of F in u and convexity in p the righthand side of this inequality remains uniformly bounded in a neighborhood of (u, p). Hence F (·, ·) is locally Lipschitz continous in p, locally uniformly in (u, p) ∈ IRN × IRnN . Therefore, if um → u , pm → p we have |F (um , pm ) − F (u, p) | ≤ |F (um , pm ) − F (um , p) | + |F (um , p) − F (u, p) | as m → ∞, ≤ c|pm − p| + o(1) → 0 where o(1) → 0 as m → ∞, as desired. (4◦) In the scalar case (N = 1), if F is C 2 for example, the existence of a minimizer u for E implies that the Legendre condition n

Fpα pβ (x, u, p) ξα ξβ ≥ 0,

for all ξ ∈ IRn

α,β=1

holds at all points (x, u = u(x), p = ∇u(x)), see for instance Giaquinta [1; p. 11 f.]. This condition in turn implies the convexity of F in p. The situation is quite diﬀerent in the vector-valued case N > 1. In this case, in general only the Legendre-Hadamard condition N i,j=1 α,β=1

Fpi pj (x, u, p)ξα ξβ η i η j ≥ 0 , α β

for all ξ ∈ IRn , η ∈ IRN

will hold at a minimizer, which is much weaker then convexity (Giaquinta [1; p. 12]). In fact, in Section 3 below we shall see how, under certain additional structure conditions on F , the convexity assumption in Theorem 1.6 can be weakened in the vector-valued case.

2. Constraints Applying the direct methods often involves a delicate interplay between the functional E, the space of admissible functions M , and the topology on M . In this section we will see how, by means of imposing constraints on admissible functions and/or by a suitable modiﬁcation of the variational problem, the direct methods can be successfully employed also in situations where their use seems highly unlikely at ﬁrst. Note that we will not consider constraints that are dictated by the problems themselves, such as physical restrictions on the response of a mechanical system. Constraints of this type in general lead to variational inequalities, and we refer to Kinderlehrer-Stampacchia [1] for a comprehensive introduction to this ﬁeld. Instead, we will show how certain variational problems can be solved

14

Chapter I. The Direct Methods in the Calculus of Variations

by adding virtual – that is, purely technical – constraints to the conditions deﬁning the admissible set, thus singling out distinguished solutions. Semilinear Elliptic Boundary Value Problems We start by deriving the existence of positive solutions to non-coercive, semilinear elliptic boundary value problems by a constrained minimization method. Such problems are motivated by studies of ﬂame propagation (see for example Gel’fand [1; (15.5), p. 357]) or arise in the context of the Yamabe problem (see Section III.4). Let Ω be a smooth, bounded domain in IRn , and let p > 2. If n ≥ 3 we also 2n . For λ ∈ IR consider the assume that p satisﬁes the condition p < 2∗ = n−2 problem (2.1) (2.2) (2.3)

−Δu + λu = u|u|p−2 u>0 u=0

in Ω , in Ω , on ∂Ω .

Also let 0 < λ1 < λ2 ≤ λ3 ≤ . . . denote the eigenvalues of the operator −Δ on H01,2 (Ω). Then we have the following result: 2.1 Theorem. For any λ > −λ1 there exists a positive solution u ∈ C 2 (Ω) ∩ C 0 (Ω) to problem (2.1)–(2.3). Proof. Observe that Equation (2.1) is the Euler-Lagrange equation of the functional 1 1 2 2 ˜ |∇u| + λ|u| dx − E(u) = |u|p dx 2 Ω p Ω on H01,2 (Ω), which is neither bounded from above nor from below on this space. However, using the homogeneity of (2.1) a solution of problem (2.1)– (2.3) can also be obtained by solving a constrained minimization problem for the functional 1 |∇u|2 + λ|u|2 dx E(u) = 2 Ω on the Hilbert space H01,2 (Ω) , restricted to the set |u|p dx = 1} . M = {u ∈ H01,2 (Ω) ; Ω

We verify that E : M → IR satisﬁes the hypotheses of Theorem 1.2. By the Rellich-Kondrakov theorem the injection H01,2 (Ω) → Lp (Ω) is completely continuous for p < 2∗ , if n ≥ 3, respectively for any p < ∞, if n = 1, 2; see Theorem A.5 of Appendix A. Hence M is weakly closed in H01,2 (Ω).

2. Constraints

15

Recall the Rayleigh-Ritz characterization

|∇u|2 dx

Ω inf (2.4) λ1 = 1,2 |u|2 dx u∈H (Ω) Ω 0 u=0

of the smallest Dirichlet eigenvalue. This gives the estimate (2.5)

E(u) ≥

1 λ u2H 1,2 . min 1, 1 + 0 2 λ1

From this, coerciveness of E for λ > −λ1 is immediate. Weak lower semi-continuity of E follows from weak lower semi-continuity of the norm in H01,2 (Ω) and the Rellich-Kondrakov theorem. By Theorem 1.2 therefore E attains its inﬁmum at a point u in M. Note that since E(u) = E(|u|) we may assume that u ≥ 0. To derive the variational equation for E ﬁrst note that E is continuously Fr´echet-diﬀerentiable in H01,2 (Ω) with v, DE(u) = ∇u∇v + λuv dx . Ω

Moreover, letting

|u|p dx − 1 ,

G(u) = Ω

for p ≤ 2∗ also G : H01,2 (Ω) → IR is continuously Fr´echet-diﬀerentiable with v, DG(u) = p u|u|p−2 v dx . Ω

In particular, at any point u ∈ M

|u|p dx = p = 0 ,

u, DG(u) = p Ω

and by the implicit function theorem the set M = G−1 (0) is a C 1 -submanifold of H01,2 (Ω). Now, by the Lagrange multiplier rule, there exists a parameter μ ∈ IR such that v, (DE(u) − μDG(u)) = ∇u∇v + λuv − μu|u|p−2 v dx Ω

= 0,

for all v ∈ H01,2 (Ω) .

Inserting v = u into this equation yields that |∇u|2 + λ|u|2 dx = μ |u|p dx = μ . 2E(u) = Ω

Ω

16

Chapter I. The Direct Methods in the Calculus of Variations

Since u ∈ M cannot vanish identically, from (2.5) we infer that μ > 0. Scaling 1 with a suitable power of μ, we obtain a weak solution u = μ p−2 · u ∈ H01,2 (Ω) of (2.1), (2.3) in the sense that ∇u∇v + λuv − u|u|p−2 v dx = 0 , for all v ∈ H01,2 (Ω) . (2.6) Ω

Moreover, (2.2) holds in the weak sense u ≥ 0, u = 0. To ﬁnish the proof we use the regularity result Lemma B.3 of Appendix B and the observations following it to obtain that u ∈ C 2 (Ω). Finally, by the strong maximum principle u > 0 in Ω ; see Theorem B.4. Observe that, at least for the kind of nonlinear problems considered here, by Lemma B.3 of Appendix B the regularity theory is taken care of and in the following we may concentrate on proving existence of (weak) solutions. However, additional structure conditions may imply further useful properties of suitable solutions. An example is symmetry. 2.2 Symmetry. By a result of Gidas-Ni-Nirenberg [1; Theorem 2.1, p. 216, and Theorem 1, p. 209], if Ω is convex and symmetric with respect to a hyperplane, say x1 = 0, any positive solution u of (2.1), (2.3) is even in x1 , that ∂u < 0 at any point is, u(x1 , x ) = u(−x1 , x ) for all x = (x1 , x ) ∈ Ω, and ∂x 1 x = (x1 , x ) ∈ Ω with x1 > 0. In particular, if Ω is a ball, any positive solution u is radially symmetric. The proof of this result uses a variant of the Alexandrov-Hopf reﬂection principle and the maximum principle. This method lends itself to numerous applications in many diﬀerent contexts; in Section III.4 we shall see that it is even possible to derive a-priori bounds from this method in the setting of a parabolic equation on the sphere. Perron’s Method in a Variational Guise In the previous example the constraint built into the deﬁnition of M had the ˜ M coercive. Moreover, this eﬀect of making the restricted functional E = E| constraint only changed the Euler-Lagrange equations by a factor which could be scaled away using the homogeneity of the right-hand side of (2.1). In the next application we will see that sometimes also inequality constraints can be imposed without changing the Euler-Lagrange equations at a minimizer. 2.3 Weak sub- and super-solutions. Suppose Ω is a smooth, bounded domain in IRn , and let g : Ω × IR → IR be a Carath´eodory function with the property that |g(x, u)| ≤ C(R) for any R > 0 and all u such that |u(x)| ≤ R almost everywhere. Given u0 ∈ H01,2 (Ω), we then consider the equation (2.7)

−Δu = g(·, u)

(2.8)

u = u0

in Ω , on ∂Ω .

2. Constraints

17

By deﬁnition u ∈ H 1,2 (Ω) is a (weak) sub-solution to (2.7–2.8) if u ≤ u0 on ∂Ω and ∇u∇ϕ dx − g( · , u)ϕ dx ≤ 0 for all ϕ ∈ C0∞ (Ω) , ϕ ≥ 0 . Ω

Ω

Similarly u ∈ H (Ω) is a (weak) super-solution to (2.7–2.8) if in the above the reverse inequalities hold. 1,2

2.4 Theorem. Suppose u ∈ H 1,2 (Ω) is a sub-solution while u ∈ H 1,2 (Ω) is a super-solution to problem (2.7–2.8) and assume that with constants c, c ∈ IR there holds −∞ < c ≤ u ≤ u ≤ c < ∞, almost everywhere in Ω. Then there exists a weak solution u ∈ H 1,2 (Ω) of (2.7–2.8), satisfying the condition u ≤ u ≤ u almost everywhere in Ω. With no loss of generality we may assume u0 = 0. Let G(x, u) =

Proof. u g(x, v) dv denote a primitive of g. Note that (2.7–2.8) formally are the 0 Euler-Lagrange equations of the functional 1 |∇u|2 dx − G(x, u) dx . E(u) = 2 Ω Ω However, our assumptions do not allow the conclusion that E is ﬁnite or even diﬀerentiable on V := H01,2 (Ω) – the smallest space where we have any chance of verifying coerciveness. Instead we restrict E to M = u ∈ H01,2 (Ω) ; u ≤ u ≤ u almost everywhere . Since u, u ∈ L∞ by assumption, also M ⊂ L∞ and G x, u(x) ≤ c for all u ∈ M and almost every x ∈ Ω. Now we can verify the hypotheses of Theorem 1.2: Clearly, V = H01,2 (Ω) is reﬂexive. Moreover, M is closed and convex, hence weakly closed. Since M is essentially bounded, our functional E(u) ≥ 12 u2H 1,2 (Ω) − c is coercive on 0 M . Finally, to see that E is weakly lower semi-continuous on M , it suﬃces to show that G(x, um ) dx → G(x, u) dx Ω

H01,2 (Ω),

Ω

where um , u ∈ M . But – passing to a subif um u weakly in sequence, if necessary – we may assume that um → u pointwise almost everywhere; moreover, |G x, um (x) | ≤ c uniformly. Hence we may appeal to Lebesgue’s theorem on dominated convergence. From Theorem 1.2 we infer the existence of a relative minimizer u ∈ M . To see that u weakly solves (2.7), for ϕ ∈ C0∞ (Ω) and ε > 0 let vε = min u, max{u, u + εϕ} = u + εϕ − ϕε + ϕε ∈ M with ϕε = max 0, u + εϕ − u ≥ 0 , ϕε = max 0, u − (u + εϕ) ≥ 0 .

18

Chapter I. The Direct Methods in the Calculus of Variations

Note that ϕε , ϕε ∈ H01,2 ∩ L∞ (Ω). E is diﬀerentiable in direction vε − u . Since u minimizes E in M we have 0 ≤ (vε − u), DE(u) = εϕ, DE(u) − ϕε , DE(u) + ϕε , DE(u) , so that

1 ε ϕ , DE(u) − ϕε , DE(u) . ε Now, since u is a supersolution to (2.7), we have ϕ, DE(u) ≥

ϕε , DE(u) = ϕε , DE(u) + ϕε , DE(u) − DE(u) ≥ ϕε , DE(u) − DE(u) ∇(u − u)∇(u + εϕ − u) = Ωε − g(x, u) − g(x, u) (u + εϕ − u) dx g(x, u) − g(x, u) |ϕ| dx , ∇(u − u)∇ϕ dx − ε ≥ε Ωε

Ωε

where Ω ε = x ∈ Ω ; u(x) + εϕ(x) ≥ u(x) > u(x) . Note that Ln (Ω ε ) → 0 as ε → 0. Hence by absolute continuity of the Lebesgue integral we obtain that ϕε , DE(u) ≥ o(ε) , where o(ε)/ε → 0 as ε → 0. Similarly, we conclude that ϕε , DE(u) ≤ o(ε) , whence

ϕ, DE(u) ≥ 0

for all ϕ ∈ C0∞ (Ω). Reversing the sign of ϕ and since C0∞ (Ω) is dense in H01,2 (Ω) we ﬁnally see that DE(u) = 0, as claimed. 2.5 A special case. Let Ω be a smooth bounded domain in IRn , n ≥ 3, and let g(x, u) = k(x)u − u|u|p−2 ,

(2.9) where p =

2n n−2 ,

and where k is a continuous function such that 1 ≤ k(x) ≤ K < ∞

uniformly in Ω. Suppose u0 ∈ C 1 (Ω) satisﬁes u0 ≥ 1 on ∂Ω. Then u ≡ 1 is a sub-solution while u ≡ c for large c > 1 is a super-solution to Equations (2.7)–(2.8). Consequently, (2.7)–(2.8) admits a solution u ≥ 1. 2.6 Remark. The sub-super-solution method can also be applied to equations on manifolds. In the context of the Yamabe problem it has been used by Loewner-Nirenberg [1] and Kazdan-Warner [1]; see Section III.4. The nonlinear term in this case is precisely (2.9).

2. Constraints

19

The Classical Plateau Problem One of the great successes of the direct methods in the calculus of variations was the solution of Plateau’s problem for minimal surfaces. Let Γ be a smooth Jordan curve in IR3 . From his famous experiments with soap ﬁlms Plateau became convinced that any such curve is spanned by a (not necessarily unique) surface of least area.

Fig. 2.1. Minimal surfaces of various topological types (disk, M¨ obius band, annulus, torus)

In the classical mathematical model the topological type of the surface is speciﬁed to be that of the disk Ω = {z = (x, y) ; x2 + y 2 < 1} . A naive approach to Plateau’s conjecture would be to attempt to minimize the area t det(∇u ∇u) dz = |ux |2 |uy |2 − (ux · uy )2 dz A(u) = Ω

Ω

among “surfaces” u ∈ H 1,2 ∩ C 0 (Ω, IR ) satisfying the Plateau boundary condition 3

(2.10)

u|

∂Ω

: ∂Ω → Γ

is a (weakly) monotone parametrization preserving the given orientation of Γ .

However, A is invariant under arbitrary changes of parameter. Hence there is no chance of achieving bounded compactness in the original variational problem and some work was necessary in order to recast this problem in a way which is accessible by direct methods. Without entering into details let us brieﬂy report the main ideas. It had already beeen observed by Lagrange that if a (smooth) surface S is (locally) area-minimizing for ﬁxed boundary Γ , necessarily the mean curvature of S vanishes. In isothermal coordinates u(x, y) on S this amounts to the equation (2.11)

Δu = 0 .

20

Chapter I. The Direct Methods in the Calculus of Variations

(See Nitsche [2] or Osserman [1].) Moreover, our choice of parameter implies the conformality relations (2.12)

|ux |2 − |uy |2 = 0 = ux · uy in Ω ,

in addition to the Plateau boundary condition (2.10). We now take equations (2.10)–(2.12) as a deﬁnition for a minimal surface spanning Γ . 2.7 The variational problem. In their 1930 break-through papers, Douglas [1] and Rad´ o [1] ingeniously proposed to solve (2.10)–(2.12) by minimizing Dirichlet’s integral 1 1 |∇u|2 dz = (|ux |2 + |uy |2 ) dz E(u) = 2 Ω 2 Ω over the class C(Γ ) = {u ∈ H 1,2 (Ω, IR3 ) ; u

∂Ω

∈ C 0 (∂Ω, IR3 ) satisﬁes (2.10)} .

It is easy to see that E(u) ≥ A(u) , and equality holds if and only if u is conformal. Actually, we have inf A(u) = inf E(u) . C(Γ )

C(Γ )

This can be derived for instance from Morrey’s “ε-conformality lemma” (Morrey [2; Theorem 1.2]). In Struwe [18; Appendix A] also a direct (variational) proof is given. Thus, a minimizer of E also will minimize A – hence it will satisfy (2.12) and solve the original minimization problem. The solution of Plateau’s problem is therefore reduced to the following theorem. 2.8 Theorem. For any C 1 -embedded curve Γ there exists a minimizer u of Dirichlet’s integral E in C(Γ ). Note that C(Γ ) = ∅ if Γ ∈ C 1 . (Actually it suﬃces to assume that Γ is a rectiﬁable Jordan curve; see Douglas [1], Rad´ o [1].) To show Theorem 2.8, observe that in replacing A by E we have succeeded in reducing the symmetries of the problem drastically. However, E is still conformally invariant, that is E(u) = E(u ◦ g) for all g ∈ G, where

a+z ; a ∈ C, |a| < 1, 0 ≤ φ < 2π G = g : z → g(z) = eiφ 1 − az

2. Constraints

21

denotes the conformal group of M¨ obius transformations of the disc, viewed as a subset of C. The action of G is non-compact in the sense that for any u ∈ C(Γ ) the orbit {u ◦ g ; g ∈ G} weakly accumulates also at constant functions; see for instance Struwe [17; Lemma I.4.1] for a detailed proof. Hence C(Γ ) cannot be weakly closed in H 1,2 (Ω, IR3 ) and Theorem 1.2 cannot yet be applied. Fortunately, we can also get rid of conformal invariance of E. Note that for any oriented triple eiφ1 , eiφ2 , eiφ3 ∈ ∂Ω , 0 ≤ φ1 < φ2 < φ3 < 2π there 2πik exists a unique g ∈ G such that g e 3 = eiφk , k = 1, 2, 3. Fix a parametrization γ of Γ (we may assume that γ is a C 1 -diﬀeomorphism γ : ∂Ω → Γ ) and let 2πik 2πik = γ e 3 , k = 1, 2, 3 , C ∗ (Γ ) = u ∈ C(Γ ) ; u e 3 endowed with the H 1,2 -topology. This is our space of admissible functions, normalized with respect to G. Note that for any u ∈ C(Γ ) there is g ∈ G such that u ◦ g ∈ C ∗ (Γ ). The following result now is a consequence of the classical “CourantLebesgue lemma”. 2.9 Lemma. The set C ∗ (Γ ) is weakly closed in H 1,2 .

Fig. 2.2.

Proof. The proof in a subtle way uses a convexity argument as in the preceding example. To present this argument explicitly we use the ﬁxed parametrization γ to associate with any u ∈ C ∗ (Γ ) a continuous map ξ : IR → IR, such that γ eiξ(φ) = u eiφ , ξ(0) = 0 . By (2.10) the functions ξ obtained in this are continuous, monotone manner 2πk , for all k ∈ ZZ by our = and ξ − id is 2π-periodic; moreover, ξ 2πk 3 3 three-point normalization.

22

Chapter I. The Direct Methods in the Calculus of Variations

Now let M = ξ :IR → IR ; ξ is continuous and monotone, 2πk 2πk = , for all φ ∈ IR, k ∈ ZZ . ξ(φ + 2π) = ξ(φ) + 2π, ξ 3 3 Note that M is convex. Let (um ) be a sequence in C ∗ (Γ ) with associated functions ξm ∈ M, and suppose um u weakly in H 1,2 (Ω). Since each ξm is monotone and satisﬁes the estimate 0 ≤ ξm (φ) ≤ 2π, for all φ ∈ [0, 2π], the family (ξm ) is bounded in BV [0, 2π] . Hence (a subsequence) ξm → ξ almost everywhere on [0, 2π] and therefore – by periodicity – almost everywhere on IR, = 2πk where ξ is monotone, ξ − id is 2π-periodic, and ξ satisﬁes ξ 2πk 3 3 , for all k ∈ ZZ. Now if ξ is continuous, it follows from monotonicity that ξm → ξ uniformly. Thus, by continuity of γ, also um converges uniformly to u on ∂Ω, and it follows that u| ∂Ω is continuous and satisﬁes (2.10). That is, u ∈ C ∗ (Γ ), and the proof is complete in this case. In order to exclude the remaining case, assume by contradiction that ξ is π discontinuous at some point φ0 . We choose k ∈ ZZ such that |φ0 − 2πk 3 | ≤ 3 2π(k−1) 2π(k+1) =: I0 . By monotonicity, for almost every φ1 , φ2 ∈ I0 and let , 3 3 such that φ1 < φ0 < φ2 we have 2π(k − 1) ≤ lim ξm (φ1 ) = ξ(φ1 ) ≤ lim ξ(φ) m→∞ 3 φ→φ− 0 < lim ξ(φ) ≤ ξ(φ2 ) = lim ξm (φ2 ) ≤ φ→φ+ 0

m→∞

2π(k + 1) . 3

For such φ1 , φ2 ∈ I0 denote I1 = {φ ∈ I0 ; φ ≤ φ1 }, I2 = {φ ∈ I0 ; φ ≥ φ2 }. Then by monotonicity of ξm and using the fact that γ is a diﬀeomorphism we obtain lim sup |γ eiξm (φ) − γ eiξm (ψ) | inf φ∈I1 , ψ∈I2 m→∞ |γ eiξ(φ) − γ eiξ(ψ) | > 0 . ≥ inf φ,ψ∈I0 ,φ<φ0 <ψ

In particular, there exists ε > 0 independent of φ1 , φ2 such that (2.13)

|um (eiφ ) − um (eiψ )| ≥ ε > 0

for all φ ∈ I1 , ψ ∈ I2 if m ≥ m0 (φ1 , φ2 ) is suﬃciently large. Now let z0 = eiφ0 and for ρ > 0 denote Uρ = {z ∈ Ω ; |z − z0 | < ρ}, Cρ = {z ∈ Ω ; |z − z0 | = ρ} . Note that for all ρ < 1 any point z = eiφ ∈ Cρ ∩ ∂Ω satisﬁes φ ∈ I0 . Following Courant [1; p. 103], we will use uniform boundedness of (um ) in H 1,2 to show that for suitable numbers ρ0 ∈]0, 1[, ρm ∈ [ρ20 , ρ0 ] the oscillation of um on Cρm can be made arbitrarily small, uniformly in m ∈ IN.

2. Constraints

23

First note that by Fubini’s theorem, if we denote arc length on Cρ by s, from the estimate

1

|∇um | dz ≥

∞>c≥

2

0

Ω

we obtain that

| Cρ

∂ um |2 ds dρ ∂s

| Cρ

∂ um |2 ds < ∞ ∂s

for almost every ρ < 1 and all m ∈ IN. Choosing ρ0 < 1 we may reﬁne this estimate as follows:

|∇um | dz ≥ 2

Ω

ρ0

ρ

∂ | um |2 ds ∂s Cρ

ρ20

dρ ρ

∂ 2 | um | ds . ρ Cρ ∂s

≥ | log ρ0 | ess inf 2

ρ0 ≤ρ≤ρ0

Suppose ρm ∈ [ρ20 , ρ0 ] is such that ρm Cρm

∂ | um |2 ds ≤ 2 ess inf ∂s ρ20 ≤ρ≤ρ0

and denote

ρ

∂ | um |2 ds ∂s Cρ

|∇um |2 dz < ∞ .

C = sup m∈IN

Ω

Fix zj = eiφj , j = 1, 2, the points of intersection of Cρ20 with ∂Ω, φ1 < φ0 < φ2 . m m Also denote zjm = eiφj , j = 1, 2, with φm 1 < φ0 < φ2 the points of intersection of Cρm with ∂Ω. m Then φm older’s inequality 1 ∈ I1 , φ2 ∈ I2 while by H¨ |um (z1m ) (2.14)

−

um (z2m )|2

≤ Cρm

∂ | um | ds ∂s

≤ πρm

| Cρm

2

∂ 2πC um |2 ds ≤ <ε ∂s | log ρ0 |

if ρ0 > 0 is suﬃciently small. This estimate being uniform in m, for large m ≥ m0 (φ1 , φ2 ) we obtain a contradiction to (2.13) and the proof is complete.

24

Chapter I. The Direct Methods in the Calculus of Variations

Remark. From (2.14), monotonicity (2.10), the assumption that Γ is a Jordan curve – that is, a homeomorphic image of the circle S 1 – and the three-point condition, also a direct proof of equi-continuity of the sub-level sets of E in C ∗ (Γ ) can be given; see Courant [1; Lemma 3.2, p. 103] or Struwe [17; Lemma I.4.3]. Moreover, note that (2.14) implies a uniform estimate for the modulus of continuity of a function u ∈ H01,2 (Ω) on a sequence of concentric circular arcs around any ﬁxed center z0 ∈ Ω in terms of its Dirichlet integral. This observation was used by Lebesgue [1] to obtain an equi-continuous minimizing sequence for Dirichlet’s integral in his solution of the classical Dirichlet problem; see also Section 5.7. Proof of Theorem 2.8. By (2.10) and the generalized Poincar´e inequality Theorem A.9 of Appendix A, for u ∈ C ∗ (Γ ) there holds

|u|2 dz ≤ c Ω

|∇u|2 dz + Ω

|u|2 do ≤ cE(u) + c(Γ ).

∂Ω

Thus E is coercive on M = C ∗ (Γ ) in H 1,2 (Ω; IR3 ). Moreover E is weakly lower semi-continuous on H 1,2 (Ω; IR3 ). By Theorem 1.2 and Lemma 2.9 therefore the functional E achieves its inﬁmum in C ∗ (Γ ), which by conformal invariance equals that in C(Γ ). 2.10 Regularity. As in the preceding examples we may ask what regularity properties the minimizer u and the parametrized surface possess. We note a few results: o [1]). (1◦) u ∂Ω is strictly monotone (Douglas [1], Rad´ ◦ m,α (2 ) If Γ ∈ C , m ≥ 1, 0 < α < 1, then also u ∈ C m,α (Ω, IR3 ); this result is due to Hildebrandt [1] (for m ≥ 4) with later improvements by Nitsche [1] . While remarks (1◦ ), (2◦ ) apply to arbitrary solutions of the Plateau problem (2.10)–(2.12), minimality is crucial for the next observations concerning the geometric regularity of the parametrized solution surface. (3◦) A minimizer u ∈ C ∗ (Γ ) parametrizes an immersed minimal surface S = u(Ω) ⊂ IR3 ; see Osserman [2], Gulliver [1], Alt [1], Gulliver-Osserman-Royden [1]; if Γ is anlytic, S is immersed up to the boundary; see Gulliver-Lesley [1]. If Γ is extreme, that is Γ ⊂ ∂K where K ⊂ IR3 is convex, S is embedded; see Meeks-Yau [1]. Existence of embedded minimal surfaces bounded by extreme curves independently was obtained by Tomi-Tromba [1] and Almgren-Simon [1]. 2.11 Note. In the history of the calculus of variations it seems that Plateau’s problem has played a very prominent role. Important developments in the general stream of ideas often were prompted by insights gained from the study of minimal surfaces. As an example, consider the classical mountain pass lemma (see also Chapter II.1) which was used by Courant [1; Chapter VI.6–7] to establish the existence of unstable minimal surfaces, previously obtained by

3. Compensated Compactness

25

Morse-Tompkins [1] and Shiﬀman [1] by a less direct, topological reasoning. However, since this material has been covered very extensively elsewhere (see Struwe [18]), here we will conﬁne ourselves to the above remarks. For an introduction to Plateau’s problem and minimal surfaces, see for instance, Osserman [1], or consult the encyclopedic book by Nitsche [2]. A truely remarkable – popular and profound – book on the subject is available by Hildebrandt-Tromba [1].

3. Compensated Compactness As noted in Remark 1.8, it is conceivable that in the vector-valued case lower semi-continuity results may hold true under a weaker convexity assumption than in Theorem 1.6, provided suitable structure conditions are satisﬁed by the functional in variation. Weakening the convexity hypothesis is necessary, for instance, in dealing with problems

arising in 3-dimensional elasticity, where we encounter energy functionals Ω W (∇u) dx with a stored energy function W depending on the determinant, the minors and the eigenvalues of the deformation gradient ∇u. Since inﬁnite volume distortion for elastic materials will aﬀord an inﬁnite amount of energy, it is natural to suppose that W → ∞ if either det(∇u) → 0 or det(∇u) → ∞; hence W cannot be convex in ∇u. However, there is a large class of materials that can be described by polyconvex stored energy functions, which are of the form W (∇u) = f (subdeterminants of ∇u), where f is convex in each of its variables. John Ball [1] was the ﬁrst to see that lower semi-continuity results will hold for such functionals. The diﬃculty, of course, lies in proving, for instance, weak convergence det(∇um ) det(∇u) for a sequence um u weakly in H 1,3 (Ω, IR3 ). Questions of this type had been investigated by Reshetnyak [1], [2]. A general frame for studying such problems is provided by the compensated compactness scheme of Murat and Tartar. The basic principle of the compensated compactness method is given in the following lemma; see Tartar [2; p. 270 f.]. 3.1 The compensated compactness lemma. Let Ω be a domain in IRn and suppose that in L2 (Ω; IRN ). (1◦) um = u1m , . . . , uN m u weakly j ∂um −1 L (2◦) The set j,k ajk ∂xk ; m ∈ IN is relatively compact in Hloc (Ω; IR ) for a set of vectors ajk ∈ IRL ; 1 ≤ j ≤ N, 1 ≤ k ≤ n. Let ajk λj ξk = 0 for some ξ ∈ IRn \ {0} Λ = λ ∈ IRN ; j,k

26

Chapter I. The Direct Methods in the Calculus of Variations

and let Q be a (real) quadratic form such that Q(λ) ≥ 0 for all ∗ λ ∈ Λ. Regarding Q(um) ∈ L1(Ω) as Radon measures Q(um )dx ∈ C 0 (Ω ) , we may assume that Q(um ) converges weak∗ , locally. Then on any Ω ⊂⊂ Ω we have weak∗ − lim Q um ≥ Q(u) m→∞

in the sense of measures. In particular, if Q(λ) = 0 for all λ ∈ Λ, then weak∗ − lim Q um = Q(u) m→∞

locally, in the sense of measures. Proof. Choose ϕ ∈ C 0 (Ω) with compact support and such that 0 ≤ ϕ ≤ 1. We must show that 2 Q um ϕ dx ≥ Q(u)ϕ2 dx . lim inf m→∞

Ω

Ω

By translation we may assume that u = 0. Moreover, upon replacing um by um ϕ we may assume that the supports of um lie in a ﬁxed cube K ⊂⊂ Ω and that ϕ ≡ 1 on K. By translation and scaling, moreover, we can obtain K = [0, 2π]n. Let N μm,α eiα·x , μm,α = μ1m,α , . . . , μN , um (x) = m,α ∈ C α∈ZZn

be the Fourier expansion for um . Since um is real, we have μm,α = μm,−α . The assertion then is equivalent to showing that 1 dx = lim inf + Q Im μ ≥0. lim inf Q u Q Re μ m m,α m,α m→∞ (2π)n K m→∞ n α∈ZZ

By weak convergence um 0 in L2 we infer that α∈ZZn |μm,α |2 ≤ c < ∞ and μm,α → 0 as m → ∞, uniformly on bounded sets of indices α. Moreover, by (2◦ ) the set ⎫ ⎧ ⎬ ⎨ ajk μjm,α αk ei α·x ; m ∈ IN ⎭ ⎩ n α∈ZZ

j,k

is relatively compact in H −1 , which implies that 2 j,k ajk μjm,α αk α∈ZZ n |α|≥α0

1 + |α|2

→ 0,

3. Compensated Compactness

27

as α0 → ∞, uniformly in m ∈ IN. But this means that μm,α can be decomposed 2 μm,α = λm,α + νm,α with Re λm,α , Im λm,α ∈ Λ and |α|≥α0 νm,α → 0 as αo → ∞, uniformly in m ∈ IN. Indeed, for any α ∈ ZZn , m ∈ IN decompose μm,α = λm,α + νm,α , where Re λm,α , Im λm,α ∈ Λ and |νm,α |2 is minimal among all decompositions of this kind. We claim that for any ε > 0 there exist constants C = C(ε), α0 = α0 (ε) such that for |α| ≥ α0 we can bound (3.1)

2 αk j + εμm,α 2 |νm,α | ≤ C ajk μm,α 1 + |α|2 2

j,k

uniformly in m ∈ IN. Otherwise there exists ε > 0 and a sequence α = α(l), l ∈ IN, with |α(l)| ≥ l, and m = m(l) such that for all l there holds (3.2)

2 αk 2 2 + ε |μm,α |2 . ajk μjm,α |μm,α | ≥ |νm,α | ≥ l 2 1 + |α| j,k

(The ﬁrst inequality follows from the choice of νm,α above.) Let ξ(l), η(l) be the unit vectors ξ(l) =

α(l) 1+

|α(l)|2

μm(l),α(l) ∈ S N−1 , ∈ S n−1 , η(l) = μm(l),α(l)

and denote by A(l): IRN → IRL the linear map η →

ajk ηj ξk (l) .

j,k

We may assume that ξ(l) → ξ and A(l) → A as l → ∞. Likewise, we may suppose that η(l) → η. Passing to the limit in (3.2) it follows that η ∈ ker A; that is, η ∈ Λ. Projecting μm,α onto ker A for all α = α(l), m = m(l) we ˜ m,α + ν˜m,α with Re λ ˜ m,α , Im λ ˜ m,α ∈ hence obtain a decomposition μm,α = λ ker A ⊂ Λ and ν˜m,α 2 ≤ C Aμm,α 2 ≤ C A(l)μm,α 2 + o(1)μm,α 2 2 α k j + o(1)μm,α 2 , =C ajk μm,α 2 1 + |α| j,k 2 2 where o(1) → 0 (l → ∞). But by deﬁntion νm,α ≤ ν˜m,α , and we obtain the desired contradiction to assumption (3.2). Hence for any ε > 0, any α0 ≥ α0 (ε):

28

lim inf m→∞

Chapter I. The Direct Methods in the Calculus of Variations

Q Re μm,α + Q Im μm,α α∈ZZn

= lim inf m→∞

≥ lim inf m→∞

Q Re μm,α + Q Im μm,α |α|≥α0

2 Q Re λm,α + Q Im λm,α − cεμm,α + o(1)

|α|≥α0

≥ o(1) − cε , where o(1) → 0 as α0 → ∞. Thus, the assertion of the lemma follows as we ﬁrst let α0 → ∞ and then ε → 0. As an application we mention the following well-known result. 3.2 The Div-Curl Lemma. Suppose um u, vm v weakly in L2 (Ω; IR3 ) on a domain Ω ⊂ IR3 while the sequences (div um ) and (curl vm ) are relatively compact in H −1 (Ω). Then for any ϕ ∈ C0∞ (Ω) we have um · vm ϕ dx → u · vϕ dx Ω

Ω

as m → ∞. Proof. Let wm = (um , vm ) ∈ L2 (Ω; IR6 ), and determine coeﬃcients ajk ∈ IR4 ∂w j such that jk ajk ∂xm = (div um , curl vm ). Let Q be the quadratic form k Q(u, v) = u·v, acting on vectors w = (u, v) ∈ IR6 . Note ajk = δjk , (εijk )1≤i≤3 where δjk = 1 if j = k, and δjk = 0 else, ε123 = 1 and εijk = −εjik = εjki . Hence Λ = λ = (μ, ν) ∈ IR6 ; ∃ξ ∈ IR3 \ {0} : (ξ · μ , ξ ∧ ν) = 0 = λ = (μ, ν) ∈ IR6 ; μ · ν = 0 , and Q ≡ 0 on Λ. Thus the assertion follows from Lemma 3.1. The div-curl Lemma 3.2 shows how additional bounds on some derivatives allow one to prove continuity of nonlinear expressions (bi-linear in the above example) under weak convergence. Phrased somewhat diﬀerently, the reason for the convergence result stated in the div-curl lemma to hold is an implicit divergence structure. This stucture can be brought out more clearly in the language of diﬀerential forms. For simplicity, we assume that all functions involved are periodic of period 1 in each variable. In this case, we may regard Ω = [0, 1]n as a fundamental domain for the ﬂat n-dimensional torus T n = IRn /ZZn . Let d, d∗ be the exterior diﬀerential and co-diﬀerential, respectively. We consider 1-forms um u in L2 , vm v

3. Compensated Compactness

29

in L2 weakly as m → ∞ with (d∗ um ), (dvm ) relatively compact in H −1 . We may assume u = 0, v = 0. By Hodge decomposition we have um = dam + d∗ bm + cm , vm = dfm + d∗ gm + hm , where cm , hm are harmonic 1-forms, dbm = dgm = 0, and with am 0, bm 0, fm 0, gm 0 weakly in H 1,2 (T n ), cm 0, hm 0 weakly in L2 (T n ). Since the space of harmonic 1-forms on T n is compact, in fact, cm → 0 and hm → 0 smoothly as m → ∞. Moreover, since the sequences Δam = d∗ um , Δgm = dvm are relatively compact in H −1 , it follows that (dam ), (d∗ gm ) are precompact in L2 , and we conclude that um = d∗ bm + o(1), vm = dfm + o(1), where o(1) → 0 in L2 as m → ∞. Moreover, using the Hodge ∗-operator and denoting by “.” contraction of forms, we have um · vm dx + o(1) = ∗(d∗ bm · dfm ) = (d ∗ bm ) ∧ dfm = d (∗bm ) ∧ dfm , thus exhibiting the asserted divergence structure. Since by Rellich’s compactness result bm → 0 strongly in L2 as m → ∞, it is now trivial to pass to the limit in the expression um · vm ϕ dx = d (∗bm ∧ dfm ϕ + o(1) Tn Tn = (−1)n (∗bm ) ∧ dfm ∧ dϕ + o(1) = o(1), Tn

where o(1) → 0 as m → ∞. A divergence structure is also the crucial ingredient in the applications that follow. Applications in Elasticity The most important applications of the compensated compactness method so far are in elasticity and hyperbolic systems, see Ball [1], [2], DiPerna [1]. DiPerna-Majda have applied compensated compactness methods to obtain the existence of weak solutions to the Euler equations for incompressible ﬂuids, see for instance DiPerna-Majda [1]. Our interest lies with the extensions of the direct methods that compensated compactness implies. Thus we will concentrate on Ball’s lower semi-continuity results for polyconvex materials in elasticity.

30

Chapter I. The Direct Methods in the Calculus of Variations

3.3 Theorem. Suppose W is a function on (3 × 3)-matrices Φ, given by W (Φ) = g(Φ, adj Φ, det Φ) where g is a convex non-negative function in the sub-determinants of Φ. Let Ω 1,3 (Ω; IR3 ). Suppose that um u weakly be a domain in IR3 and let um , u ∈ Hloc 3 in H 1,3 (Ω ; IR ) while det(∇um ) h weakly in L1 (Ω ) for all Ω ⊂⊂ Ω, where h ∈ L1loc (Ω). Then

W (∇u) dx ≤ lim inf Ω

m→∞

W (∇um ) dx . Ω

Proof. The proof of Theorem 1.6 can be carried over once we show that under the hypotheses made adj(∇um ) adj(∇u) det(∇um ) det(∇u) weakly in L1 (Ω ) for all Ω ⊂⊂ Ω. The ﬁrst assertion is a consequence of the divergence structure of the adjoint matrix Am = adj (∇um ). Indeed, if indices i, j are counted modulo 3 we have ∂ui+2 ∂ui+1 ∂ui+1 ∂ui+2 m m m m − ∂xj+1 ∂xj+2 ∂xj+1 ∂xj+2 i+2 i+2 ∂ ∂ i+1 ∂um i+1 ∂um = um − um . ∂xj+1 ∂xj+2 ∂xj+2 ∂xj+1

Aij m =

3/2 Fix Ω ⊂⊂ Ω. Note that Aij (Ω ). Hence we may asm is bounded in L ij ij 3/2 sume that Am A weakly in L (Ω ). Moreover, by Rellich’s theorem um → u in L3 (Ω ), whence um ∇um u∇u weakly in L3/2 (Ω ). By continuity of the distributional derivative with respect to weak convergence therefore ij in the sense of distributions. Finally, by uniqueness of the Aij m adj(∇u) ij distributional limit, Aij = adj(∇u) , and adj(∇um ) adj(∇u) weakly in L3/2 (Ω ), in particular weakly in L1 (Ω ), as claimed. Similarly, expanding the determinant along the ﬁrst row, we have # $ ∂u1m ∂u2m ∂u3m ∂u2m ∂u3m det(∇um ) = − ∂x1 ∂x2 ∂x3 ∂x3 ∂x2 # $ ∂u2m ∂u3m ∂u1m ∂u2m ∂u3m − − ∂x2 ∂x1 ∂x3 ∂x3 ∂x1 # $ ∂u2m ∂u3m ∂u1m ∂u2m ∂u3m − . + ∂x3 ∂x1 ∂x2 ∂x2 ∂x1

3. Compensated Compactness

31

Again, a divergence structure emerges, if we rewrite this as # 2 $ ∂ ∂u2m ∂u3m ∂um ∂u3m 1 det(∇um ) = − um ∂x1 ∂x2 ∂x3 ∂x3 ∂x2 # 2 $ ∂um ∂u3m ∂u2m ∂u3m ∂ 1 − um − ∂x2 ∂x1 ∂x3 ∂x3 ∂x1 # 2 $ ∂um ∂u3m ∂u2m ∂u3m ∂ 1 u . − + m ∂x3 ∂x1 ∂x2 ∂x2 ∂x1 Thus convergence det(∇um ) → det(∇u) in the sense of distributions follows by weak convergence in L3/2 (Ω ) of the terms in brackets [- -], proved above, strong convergence um → u in L3 (Ω ), and weak continuity of the distributional derivative. Finally, by uniqueness of the weak limit in the distribution sense, it follows that det(∇u) = h and det(∇um ) det(∇u) weakly in L1loc , as claimed.

Observe that in the language of diﬀerential forms the divergence structure of a Jacobian or its minors is even more apparent. In fact, for any smooth function u = (u1 , u2 , u3 ): Ω ⊂ IR3 → IR3 we have det(∇u) dx = ∗ det(∇u) = du1 ∧ du2 ∧ du3 , where d denotes exterior derivative and where ∗ denotes the Hodge star operator (which in this case converts a function on Ω into a 3-form). Now dd = 0, and therefore du1 ∧ du2 ∧ du3 = d(u1 du2 ∧ du3 ), which immediately implies the asserted divergence structure. Moreover, this result (and therefore Theorem 3.3) generalizes to any dimension n, for um u 1,n (Ω; IRn ) with det(∇um ) h weakly in L1loc (Ω) as m → ∞. weakly in Hloc The assumption det(∇um ) h ∈ L1loc (Ω) at ﬁrst sight may appear rather awkward. However, examples by Ball-Murat [1] show that weak H 1,3 convergence in general does not imply weak L1 -convergence of the Jacobian. This diﬃculty does not arise if we assume weak convergence in H 1,3+ε for some ε > 0. Hence, adding appropriate growth conditions on W to ensure coerciveness of the functional Ω W (∇u) dx on the space H 1,3+ε (Ω; IR3 ) for some ε > 0, from Theorem 3.3 the reader can derive existence theorems for deformations of elastic materials involving polyconvex stored energy functions. As a further reference for such results, see Ciarlet [1] or Dacorogna [1], [2]. Recently, more general results on weak continuity of determinants and corresponding existence theorems in nonlinear elasticity have been obtained by GiaquintaModica-Souˇcek [1] and M¨ uller [1], [2], [3]. The regularity theory for problems in nonlinear elasticity is still evolving. Some material can be found in the references cited above. In particular, the

32

Chapter I. The Direct Methods in the Calculus of Variations

question of cavitation of elastic materials has been studied. See for instance Giaquinta-Modica-Souˇcek [1]. Convergence Results for Nonlinear Elliptic Equations We close this section with another simple and useful example of how compensated compactness methods may be applied in a nonlinear situation. The following result is essentially “Murat’s lemma” from Tartar [2, p. 278]: 3.4 Theorem. Suppose um ∈ H01,2 (Ω) is a sequence of solutions to an elliptic equation in Ω −Δum = fm um = 0 on ∂Ω in a smooth and bounded domain Ω in IRn . Suppose um u weakly in H01,2 (Ω) while (fm ) is bounded in L1 (Ω). Then for a subsequence m → ∞ we have ∇um → ∇u in Lq (Ω) for any q < 2, and ∇um → ∇u pointwise almost everywhere. Proof. Choose p > n and let ϕm ∈ H01,p (Ω) satisfy ϕm H 1,p ≤ 1 0

(∇um − ∇u)∇ϕm dx = Ω

(∇um − ∇u)∇ϕ dx .

sup ϕ∈H01,p (Ω), ϕ

1,p ≤1 H 0

Ω

By the Calder´ on-Zygmund inequality in Lp , see Simader[1], the latter sup (∇um − ∇u)∇ϕ dx ≥ c−1 um − uH 1,q ϕ∈H01,p (Ω), ϕ

where

1 p

+

1 q

1,p ≤1 H 0

0

Ω

= 1. n

On the other hand, by Sobolev’s embedding H01,p (Ω) → C 1− p (Ω). Hence by the Arz´ela-Ascoli theorem we may assume that ϕm ϕ weakly in H01,p (Ω) and uniformly in Ω. (See Theorem A.5.) Thus (∇um − ∇u)∇ϕm dx = (∇um − ∇u)(∇ϕm − ∇ϕ) dx + o(1) Ω Ω = lim (∇um − ∇ul )(∇ϕm − ∇ϕ) dx + o(1) l→∞ Ω (fm − fl )(ϕm − ϕ) dx + o(1) = lim l→∞

Ω

≤ 2 sup fl L1 ϕm − ϕL∞ + o(1) = o(1), l∈IN

where o(1) → 0 as m → ∞.

3. Compensated Compactness

33

It follows that ∇um → ∇u in Lq0 (Ω) for some q0 ≥ 1. But then, by H¨ older’s inequality, for any q < 2 we have ∇um − ∇uLq ≤ ∇um − ∇uγL1 ∇um − ∇u1−γ L2 → 0 , where

1 q

=γ+

1−γ . 2

Results like Theorem 3.4 are needed if one wants to solve nonlinear partial diﬀerential equations −Δu = f (x, u, ∇u)

(3.3) with quadratic growth

|f (x, u, p)| ≤ c(1 + |p|2 ) by approximation methods. Assuming some uniform control on approximate solutions um of (3.3) in H 1,2 , Theorem 3.4 assures that f (x, um , ∇um ) → f (x, u, ∇u) almost everywhere, where u is the weak limit of a suitable sequence (um ). Given some further structure conditions on f , then there are various ways of passing to the limit m → ∞ in Equation (3.3); see, for instance, Frehse [2] or Evans [2]. 3.5 Example. As a model problem consider the equation (3.4) (3.5)

A(u) = −Δu + u|∇u|2 = h in Ω u=0 on ∂Ω

on a smooth and bounded domain Ω ⊂ IRn , with h ∈ L∞ (Ω). This is a special case of a problem studied by Bensoussan-Boccardo-Murat [1; Theorem 1.1, p. 350]. Note that the nonlinear term g(u, p) = u|p|2 satisﬁes the condition g(u, p)u ≥ 0 .

(3.6) Approximate g by functions

gε (u, p) =

g(u, p) ,ε > 0 , 1 + ε|g(u, p)|

satisfying |gε | ≤ 1ε and gε (u, p) · u ≥ 0 for all u, p. Now, since gε is uniformly bounded, the map H01,2 (Ω) u → gε (u, ∇u) ∈ −1 H (Ω) is compact and bounded for any ε > 0. Denote Aε (u) = −Δu + gε (u, ∇u) the perturbed operator A. By Schauder’s ﬁxed point theorem, see, for instance, Deimling [1; Theorem 8.8, p. 60], applied to the map u → (−Δ)−1 h− gε (u, ∇u) on a suﬃciently large ball in H01,2 (Ω), there is a solution uε ∈ H01,2 (Ω) of the equation Aε uε = h for any ε > 0. In addition, since gε (u, p)·u ≥ 0 we have

34

Chapter I. The Direct Methods in the Calculus of Variations

uε 2H 1,2 ≤ uε , Aε uε = uε , h ≤ uε H 1,2 hH −1 , 0

0

and (uε ) is uniformly bounded in H01,2 (Ω) for ε > 0. Moreover, since the nonlinear term gε satisﬁes 1 + |u|2 |p|2 u|p|2 ≤ gε (u, p) = 1 + ε|u| |p|2 1 + ε|u| |p|2 ≤ |p|2 + gε (u, p)u , we also deduce the uniform L1 -bound gε (uε , ∇uε )

L1

≤

uε 2H 1,2 0

uε gε (uε , ∇uε ) dx

+ Ω

= uε , Aε uε ≤ c . We may assume that a sequence (uεm ) as εm → 0 weakly converges in H01,2 (Ω) to a limit u ∈ H01,2 (Ω). By Theorem 3.4, moreover, we may assume that um = uεm converges strongly in H01,q (Ω) and that um and ∇um converge pointwise almost everywhere. To show that u weakly solves (3.4), (3.5) we now use the “Fatou lemma technique” of Frehse [2]. As a preliminary step we establish a uniform L∞ -bound for the sequence (um ). Multiply the approximate equations by um to obtain the diﬀerential inequality −Δ

|um |2 |um |2 ≤ −Δ + |∇um |2 + um gεm (um , ∇um ) 2 2 |um |2 , = hum ≤ C(δ) + δ 2

for any δ > 0. Choosing δ < λ1 , the ﬁrst eigenvalue of −Δ on H01,2 (Ω), the weak maximum principle implies a uniform bound for um in L∞ . (See Theorem B.7 and its application in Appendix B.) Next, testing the approximate equations Aεm (um) = h with ϕ = ξ exp(γum ), where ξ ∈ C0∞ (Ω) is non-negative, upon integrating by parts we obtain γ|∇um |2 + gεm (um , ∇um ) ξ exp(γum ) dx Ω ∇um ∇ξ − hξ exp(γum ) dx = 0 . + Ω

Note that on account of the growth condition |gεm (u, p)| ≤ |u| |p|2 and the uniform bound um L∞ ≤ C0 derived above, for |γ| ≥ C0 the term γ|∇um |2 + gεm (um , ∇um )

3. Compensated Compactness

35

has the same sign as γ. Moreover, this term converges pointwise almost everywhere to the same expression involving u instead of um . Hence, by Fatou’s lemma, upon passing to the limit m → ∞ we obtain ξ exp(γu), A(u) − h ≤ 0 if γ ≥ C0 , respectively ≥ 0, if γ ≤ −C0 . This holds for all non-negative ξ ∈ C0∞ (Ω) and hence also for ξ ≥ 0 belonging to H01,2 ∩ L∞ (Ω). Setting ξ = ξ0 exp(−γu) we obtain ξ0 , A(u) − h ≤ 0 and ≥ 0, for any ξ0 ≥ 0, ξ0 ∈ C0∞ (Ω). Hence u is a weak solution of (3.4), (3.5), as desired. More sophisticated variants and applications of Theorem 3.4 are given by Bensoussan-Boccardo-Murat [1], Boccardo-Murat-Puel [1], and Frehse [1], [2].

Hardy Space Methods The Hardy Hp -spaces play an important role in harmonic analysis. In the theory of partial diﬀerential equations, the space H1 is of particular interest. For instance, the Calder´ on-Zygmund theorem, asserting that for 1 < p < ∞ 2,p , ceases to be valid in any function u ∈ Lp with Δu ∈ Lp belongs to Hloc the limit case p = 1. However, the theorem remains true if we substitute L1 by the slightly smaller space H1 . While the harmonic analysts’ deﬁnition of H1 is rather unwieldy and hardly lends itself to applications in the theory of partial diﬀerential equations, recently an observation of M¨ uller [2], [3] has led to the discovery of simple criteria for composite functions to belong to H1 . In particular, it was shown by Coifman-Lions-Meyer-Semmes [1] that the Jacobian of a function u ∈ H 1,n (IRn ; IRn ) belongs to this space, and similarly for l × l-sub-determinants of ∇u for u ∈ H 1,l (IRn ; IRN ), for any l, n, N ∈ IN. As an application, we derive the assertion of Theorem 3.4 in the case that for a sequence um u weakly in H 1,2 (IRn ; IRN ) the sequence −Δum = fm is on-Zygmund theorem to bounded in H1 . In fact, by the extension of the Calder´ H1 , the sequence (∇2 um ) is bounded in H1 ⊂ L1 , and therefore, by Rellich’s compactness result, ∇um → ∇u locally in L1 as m → ∞. Boundedness of (fm ) in H1 by the result of Coifman-Lions-Meyer-Semmes [1] is guaranteed if, for instance, for each m ∈ IN the function fm is a linear combination of 2 × 2 subdeterminants of ∇F (um ), where F : IRN → IRL is a smooth, bounded function with bounded derivative. For more information about Hardy spaces and their applications, see, for instance, Torchinsky [1] or Semmes [1].

36

Chapter I. The Direct Methods in the Calculus of Variations

4. The Concentration-Compactness Principle As we have seen in our analysis of the Plateau problem, Section 2.7, a very serious complication for the direct methods to be applicable arises in the presence of non-compact group actions. If, in a terminology borrowed from physics which we will try to make more precise later, the action is a “manifest” symmetry – as in the case of the conformal group of the disk acting on Dirichlet’s integral for minimal surfaces – we may be able to eliminate the action by a suitable normalization. This was the reason for introducing the three-point condition on admissible functions for the Plateau problem in the proof of Theorem 2.8. However, if the action is “hidden”, such a normalization is not possible and there is no hope that all minimizing sequences converge to a minimizer. Even worse, the variational problem may not have a solution. For such problems, P.-L. Lions developed his concentration-compactness principle. On the basis of this principle, for many constrained minimization problems it is possible to state necessary and suﬃcient conditions for the convergence of all minimizing sequences satisfying the given constraint. These conditions involve a delicate comparison of the given functional in variation and a (family of) functionals “at inﬁnity” (on which the group action is manifest). Rather than dwell on abstract notions we prefer to give an example – a variant of problem (2.1), (2.3) – which will bring out the main ideas immediately. 4.1 Example. that

Let a: IRn → IR be a continuous function a > 0 and suppose as |x| → ∞ .

a(x) → a∞ > 0

We look for positive solutions u of the equation (4.1)

−Δu + a(x)u = u|u|p−2

in IRn ,

decaying at inﬁnity, that is (4.2)

u(x) → 0

as |x| → ∞ .

Here p > 2 may be an arbitrary number, if n = 1, 2. If n ≥ 3 we suppose that 2n . This guarantees that the imbedding p < n−2 H 1,2 (Ω) → Lp (Ω) is compact for any Ω ⊂⊂ IRn . Note that, as in the proof of Theorem 2.1, solutions of Equation (4.1) correspond to critical points of the functional 1 |∇u|2 + a(x)|u|2 dx E(u) = 2 IRn on H 1,2 (IRn ), restricted to the unit sphere

4. The Concentration-Compactness Principle

37

M = {u ∈ H 1,2 (IRn ) ;

IRn

|u|p dx = 1}

in Lp (IRn ). Moreover, if a(x) ≡ a∞ , E is invariant under translations u → ux0 (x) = u(x − x0 ) . In general, for any u ∈ H 1,2 (IRn ), after a substitution of variables 1 1 |∇u|2 + a(x + x0 )|u|2 dx → |∇u|2 + a∞ |u|2 dx E(ux0 ) = 2 IRn 2 IRn as |xo | → ∞, whence it may seem appropriate to call 1 ∞ |∇u|2 + a∞ |u|2 dx E (u) := 2 IRn the functional at inﬁnity associated with E. The following result is due to Lions [2; Theorem I.2]. 4.2 Theorem. Suppose (4.3)

I := inf E < inf E ∞ =: I ∞ , M

M

then there exists a positive solution u ∈ H 1,2 (IRn ) of Equation (4.1). Moreover, condition (4.3) is necessary and suﬃcient for the relative compactness of all minimizing sequences for E in M . Proof. Clearly, (4.3) is necessary for the relative compactness of all minimizing sequences in M . Indeed, suppose I ∞ ≤ I and let (um ) be a minimizing sequence um ), given by u ˜m = um (· + xm ), is a minimizing sequence for E ∞ . Then also (˜ for E ∞ , for any sequence (xm ) in IRn . Choosing |xm | large enough such that 1 E(˜ um ) − E ∞ (˜ um ) ≤ , m moreover, (˜ um ) is a minimizing sequence for E. In addition, we can achieve that locally in L2 , u ˜m → 0 whence (˜ um ) cannot be relatively compact. Note that this argument also proves that the weak inequality I ≤ I ∞ always holds true, regardless of the particular choice of the function a. We now show that condition (4.3) is also suﬃcient for the relative compactness of minimizing sequences. The existence of a positive solution to (4.1) then follows as in the proof of Theorem 2.1. Let (um ) be a minimizing sequence for E in M with E(um ) → I. We may assume that um u weakly in Lp (IRn ) . By continuity, a is uniformly positive on IRn . Hence we also have

38

Chapter I. The Direct Methods in the Calculus of Variations

um 2H 1,2 ≤ c E(um ) ≤ C < ∞ , and in addition we may assume that um u weakly in H 1,2 (IRn ) and pointwise almost everywhere. Let um = vm + u. Observe that the family (u|um − ϑu|p−1 ) is equi-integrable on IRn , uniformly in 0 ≤ ϑ ≤ 1. Hence by Vitali’s theorem 1 d |um |p dx − |um − u|p dx = − |um − ϑu|p dϑ dx 0 dϑ 1 (4.4) u(um − ϑu)|um − ϑu|p−2 dϑ dx =p 0

1

→p where

u(u − ϑu)|u − ϑu|p−2 dϑ dx =

|u|p dx ,

0

... dx denotes integration over IRn ; that is, p |u| dx + |vm |p dx → 1 . IRn

IRn

Similarly, we have

(4.5)

E(um ) = E(vm + u) 1 |∇u|2 + 2∇u∇vm + |∇vm |2 = 2 IRn + a(x) |u|2 + 2uvm + |vm |2 dx = E(u) + E(vm ) + ∇u∇vm + a(x)uvm dx, IRn

and the last term converges to zero by weak convergence vm = um − u 0 in H 1,2 (IRn ). Moreover, for any ε > 0, letting Ωε = {x ∈ IRn ; |a(x) − a∞ | ≥ ε} ⊂⊂ IRn , since vm → 0 locally in L2 , the integral (a(x) − a∞ )|vm |2 dx IRn

≤ε

|vm | + sup |a(x)|

|vm |2 dx

2

IRn

IRn

Ωε

≤ cε + o(1) . Here and in the following, o(1) denotes error terms such that o(1) → 0 as m → ∞. Hence this integral can be made arbitrarily small if we ﬁrst choose ε > 0 suﬃciently small and then let m ≥ m0 (ε) be suﬃciently large. That is, we have

4. The Concentration-Compactness Principle

39

E(um ) = E(u) + E ∞ (vm ) + o(1) .

By homogeneity, if we denote λ = IRn |u|p dx, E(u) = λ2/p E λ−1/p u ≥ λ2/p I, if λ > 0 , E ∞ (vm ) = (1 − λ)2/p E ∞ (1 − λ)−1/p vm ≥ (1 − λ)2/p I ∞ + o(1), if λ < 1 . Hence, for all λ ∈ [0, 1], we obtain the estimate I = E(um ) + o(1) = E(u) + E ∞ (vm ) + o(1) ≥ λ2/p I + (1 − λ)2/p I ∞ + o(1) ≥ λ2/p + (1 − λ)2/p I + o(1) . Since p > 2 this implies that λ ∈ {0, 1}. But if λ = 0, we obtain that I ≥ I ∞ + o(1) > I for large m; a contradiction. Therefore λ = 1; that is, um → u in Lp , and u ∈ M. By convexity of E, moreover, E(u) ≤ lim inf E(um ) = I , m→∞

and u minimizes E in M . Hence also E(um ) → E(u). Finally, by (4.5) um − u2H 1,2 ≤ cE(um − u) = c (E(um ) − E(u)) + o(1) → 0 , and um → u strongly in H 1,2 (IRn ). The proof is complete. Regarding |um |p dx as a measure on IRn , a systematic approach to such problems is possible via the following lemma (P.L. Lions [1; p. 115 ﬀ.]). 4.3 Concentration-Compactness Lemma I . Suppose μm is a sequence of prob ability measures on IRn : μm ≥ 0, IRn dμm = 1. There is a subsequence (μm ) such that one of the following three conditions holds: (1◦) (Compactness) There exists a sequence xm ⊂ IRn such that for any ε > 0 there is a radius R > 0 with the property that dμm ≥ 1 − ε BR (xm )

for all m. (2◦) (Vanishing) For all R > 0 there holds dμm = 0 . lim sup m→∞

x∈IRn

BR (x)

40

Chapter I. The Direct Methods in the Calculus of Variations

(3◦) (Dichotomy) There exists a number λ, 0 < λ < 1, such that for any ε > 0 there is a number R > 0 and a sequence (xm ) with the following property: Given R > R there are non-negative measures μ1m , μ2m such that 0 ≤ μ1m + μ2m ≤ μm , supp(μ1m ) ⊂ BR (xm ), supp(μ2m ) ⊂ IRn \ BR (xm ) , lim sup λ − dμ1m + (1 − λ) − dμ2m ≤ ε . n n m→∞ IR

IR

Proof. The proof is based on the notion of concentration function Q(r) = sup

dμ

x∈IRn

Br (x)

of a non-negative measure, introduced by L´evy [1]. Let Qm be the concentration functions associated with μm . Note that (Qm ) is a sequence of non-decreasing, non-negative bounded functions on [0, ∞[ with limR→∞ Qm (R) = 1. Hence, (Qm ) is locally bounded in BV on [0, ∞[ and there exists a subsequence (μm ) and a bounded, non-negative, non-decreasing function Q such that Qm (R) → Q(R)

(m → ∞) ,

for almost every R > 0. We normalize Q to be continuous from the left. Since Qm is non-decreasing, this then also implies that for every R > 0 we have Q(R) ≤ lim inf Qm (R). m→∞

Let λ = lim Q(R) . R→∞

Clearly 0 ≤ λ ≤ 1. If λ = 0, we have “vanishing”, case (2◦ ). Suppose λ = 1. Then for some R0 > 0 we have Q(R0 ) > 12 . For any m ∈ IN let xm satisfy 1 dμm + Qm (R0 ) ≤ . m BR0 (xm ) Now for 0 < ε <

1 2

ﬁx R such that Q(R) > 1 − ε > 12 and let ym satisfy 1 . dμm + Qm (R) ≤ m BR (ym )

Then, with error o(1) → 0 as m → ∞, we have dμm + dμm ≥ Qm (R0 ) + Qm (R) + o(1) BR (ym )

BR0 (xm )

≥ Q(R0 ) + Q(R) + o(1)

4. The Concentration-Compactness Principle

41

and the right-hand side is >1=

IRn

dμm

for suﬃciently large m. It follows that for such m BR (ym ) ∩ BR0 (xm ) = ∅ . That is, BR (ym ) ⊂ B2R+R0 (xm ) and hence 1−ε≤

dμm

B2R+R0 (xm )

for large m. Choosing R even larger, if necessary, we can achieve that (1◦ ) holds for all m. If 0 < λ < 1, given ε > 0 choose R and a sequence (xm ) – depending on ε and R – such that dμm > λ − ε , Qm (R) ≥ BR (xm )

if m ≥ m0 (ε). Enlarging m0 (ε), if necessary, we can also ﬁnd a sequence Rm → ∞ such that Qm (R) ≤ Qm (Rm ) < λ + ε , if m ≥ m0 (ε). Moreover, given R > R, we may assume that Rm ≥ R for all m. Now let μ1m = μm BR (xm ), the restriction of μm to BR (xm ). Similarly, deﬁne μ2m = μm IRn \ BRm (xm ) . Obviously 0 ≤ μ1m + μ2m ≤ μm , and supp(μ1m ) ⊂ BR (xm ), supp(μ2m ) ⊂ IRn \ BRm (xm ) ⊂ IRn \ BR (xm ) . Finally, for m ≥ m0 (ε) we can estimate 1 2 1 − λ − λ − + dμ dμ m m IRn IRn = λ − dμm + dμm − λ < 2ε , BRm (xm ) BR (xm ) which concludes the proof. In the context of Theorem 4.2, Lemma 4.3 may be applied to μm = |um |p dx, m ∈ IN. Dichotomy in this case is made explicit in (4.4). In view of the compactness of the embedding H 1,2 (Ω) → Lp (Ω) on bounded domains Ω for 2n the situation dealt with in Example 4.1 is referred to as the locally all p < n−2 compact case. Further complications arise in the presence of non-compact symmetry groups acting locally; for instance, in the case of conformal invariance or invariance under scaling.

42

Chapter I. The Direct Methods in the Calculus of Variations

Existence of Extremal Functions for Sobolev Embeddings A typical example is the case of Sobolev’s embedding on a (possibly unbounded) domain Ω ⊂ IRn . 4.4 Sobolev embeddings. For u ∈ C0∞ (Ω), k ≥ 1, p ≥ 1, let upDk,p = |Dα u|p dx , |α|=k

Ω

and let Dk,p (Ω) denote the completion of C0∞ (Ω) in this norm. Suppose kp < n. By Sobolev’s embedding, Dk,p (Ω) → Lq (Ω) where 1q = 1p − nk , and there exists a (maximal) constant S = S(Ω) = S(Ω, k, n, p) such that SupLq ≤ upDk,p , for all u ∈ Dk,p (Ω) .

(4.6)

Using Schwarz-symmetrization (see for instance Polya-Szeg¨ o [1; Note A.5, p. 189 ﬀ.]), for k = 1, best constants and extremal functions (on Ω = IRn ) can be computed classically, see Talenti [1]; the earliest result in this regard seems to be due to Rodemich[1]. But for k > 1 this method can no longer be applied. Using the concentration-compactness principle, however, the existence of extremal functions for Sobolev’s embedding can be established in general; see Theorem 4.9 below. First we note an important property of the embedding (4.6). 4.5 Scale invariance. By invariance of the norms in Dk,p (IRn ), respectively Lq (IRn ), under translation and scaling u → uR (x) = R−n/q u(x/R)

(4.7)

the Sobolev constant S is independent of Ω. Indeed, for any domain Ω, extending a function u ∈ C0∞ (Ω) by 0 outside Ω, we may regard C0∞ (Ω) as a subset of C0∞ (IRn ). Similarly, we may regard Dk,p (Ω) as a subset of Dk,p (IRn ). Hence we have S(Ω) = inf upDk,p ; u ∈ Dk,p (Ω), uLq = 1 ≥ S(IRn ) . Conversely, if um ∈ Dk,p (IRn ) is a minimizing sequence for S(IRn ) with um Lq = 1, by density of C0∞ (IRn ) in Dk,p (IRn ) we may assume that um ∈ C0∞ (IRn ). After translation, moreover, we have 0 ∈ Ω. Scaling with (4.7), for suﬃciently small Rm we can achieve that vm = (um )Rm ∈ C0∞ (Ω). But by invariance of · Dk,p , · Lq under (4.7) there now results S(Ω) ≤ lim inf vm pDk,p = S(IRn ) , m→∞

n

and S(Ω) = S(IR ) = S, as claimed.

4. The Concentration-Compactness Principle

43

4.6 The case k = 1. For k = 1 the Sobolev inequality has an underlying geometric meaning which allows one to analyze this case completely. Consider n , we ﬁrst the case p = 1. For u ∈ D1,1 (IRn ), or even u ∈ BV (IRn ), q1 = n−1 claim 1/q1 1 |u|q1 dx ≤ |∇u| dx , (4.8) 1/n IRn n1/q1 ωn−1 IRn where ωn−1 denotes the (n − 1)-dimensional measure of the unit sphere in IRn . Observe that equality holds if and only if u ∈ BV (IRn ) is a scalar multiple of the characteristic function of a ball in IRn . This reﬂects the fact that 1/q1 n L (Ω) 1 K(n) = = sup n 1/n Ω⊂⊂IRn P (Ω; IR ) n1/q1 ω n−1

equals the isoperimetric constant in IRn , and K(n) is achieved if and only if Ω is a ball in IRn . (The perimeter P (Ω; IRn ) was deﬁned in Theorem 1.4.) The following proof of (4.8), based on Talenti [2; p. 404], reveals the deep relation between isoperimetric inequalities and best constants for Sobolev embeddings more clearly. (See also Cianchi [1; Lemma 1].) Proof of (4.8). For u ∈ C0∞ (IRn ), and t ≥ 0 let Ω(t) = {x ∈ IRn ; |u(x)| > t} .

Then

∞

|u(x)| =

χΩ(t) (x) dt 0

for almost every x ∈ Ω, and hence by Minkowsky’s inequality ∞ uLq1 ≤ χΩ(t) Lq1 dt 0 ∞ 1/q1 Ln Ω(t) dt = 0 ∞ P Ω(t); IRn dt . ≤ K(n) 0

Finally, by the co-area formula Giusti [1; Theorem 1.23]) |∇u| dx = IRn

(see for instance Federer [1; Theorem 3.2.11] or

IRn

∇|u| dx =

∞

P Ω(t); IRn dt ,

0

and (4.8) follows. From (4.8) the general case p ≥ 1 can be derived by applying H¨ older’s np n ∞ = sq1 , where s = np−p ≥ 1. Then for u ∈ C inequality. Denote q = n−p 0 (IR ) n−p we can estimate

44

Chapter I. The Direct Methods in the Calculus of Variations

% %1/s 1/s uLq = %|u|s %Lq1 ≤ K(n) 1/s ≤ sK(n)

IRn

IRn

∇|u|s dx

1/s

1/s

|∇u| |u|s−1 dx

1/s 1/s 1−1/s ≤ sK(n) uD1,p uLq and (4.9)

uLq ≤ sK(n)uD1,p .

(We do not claim that this constant is sharp.) 4.7 Bounded domains. In contrast to the case p = 1, for p > 1 the best constant in inequality (4.9) is never achieved on any domain Ω diﬀerent from IRn ; in particular, it is never achieved on a bounded domain. Indeed, if u ∈ D1,p (Ω) achieves S = S(IRn ), a multiple of u weakly solves the equation in IRn (4.10) −∇ |∇u|p−2 ∇u = u|u|q−2 and vanishes on IRn \ Ω, which contradicts the strong maximum principle for Equation (4.10); see, for instance, Tolksdorf [1]. Of course, we suspect invariance under scaling (4.7) to be responsible for this defect. Note that for any u ∈ Dk,p (IRn ) there holds uR 0 weakly in Dk,p (IRn ) as R → 0 , while S is invariant under scaling. Hence, relative compactness of minimizing sequences cannot be expected. Observe, moreover, that for u ∈ C0∞ (IRn ) the support of uR lies in a ﬁxed compact set for all R ≤ 1. That is, we encounter a new type of loss of compactness compared to Example 4.1; we are dealing with a problem which is also locally non-compact. This is the setting for the second concentration-compactness lemma from P.L. Lions [3; Lemma I.1]. Denote |Dα u|p = |Dk u|p , |α|=k

for convenience. 4.8 Concentration-Compactness Lemma II. Let k ∈ IN, p ≥ 1, kp < n, 1q = 1 − nk . Suppose um u weakly in Dk,p (IRn ) and μm = |∇k um |p dx μ , p νm = |um |q dx ν weakly in the sense of measures where μ and ν are bounded non-negative measures on IRn . Then we have: (1◦) There exists some at most countable set J, a family {x(j) ; j ∈ J} of distinct points in IRn , and a family {ν (j) ; j ∈ J} of positive numbers such that

4. The Concentration-Compactness Principle

ν = |u|q dx +

45

ν (j) δx(j) ,

j∈J

where δx is the Dirac-mass of mass 1 concentrated at x ∈ IRn . (2◦) In addition we have μ(j) δx(j) μ ≥ |∇k u|p dx + j∈J

for some family {μ(j) ; j ∈ J}, μ(j) > 0 satisfying p/q ≤ μ(j) , S ν (j) In particular,

j∈J

for all j ∈ J .

(j) p/q < ∞. ν

Proof. Let vm = um − u ∈ Dk,p (IRn ). Then vm 0 weakly in Dk,p , and by (4.4) we have ωm := νm − |u|q dx = |um |q − |u|q dx = |um − u|q dx + o(1) = |vm |q dx + o(1) , where o(1) 0 as m → ∞. Also let λm := |∇k vm |p dx. We may assume that λm λ, while ωm ω = ν − |u|q dx weakly in the sense of measures, where λ, ω ≥ 0. Choose ξ ∈ C0∞ (IRn ). Then |ξ|q dω = lim |ξ|q dωm = lim |vm ξ|q dx IRn

m→∞

IRn

≤ S −q/p lim inf m→∞

=S

−q/p

IRn

IRn

m→∞

IRn

q/p

|∇k (vm ξ)|p dx q/p

|ξ| |∇ vm | dx p

lim inf m→∞

= S −q/p

IRn

k

p

q/p |ξ|p dλ .

Observe that by Rellich’s theorem any lower order terms like |∇l ξ||∇k−l vm | → 0 in Lp , as m → ∞. That is, there holds p/q q |ξ| dω ≤ |ξ|p dλ (4.11) S IRn

IRn

Now let {x(j) ; j ∈ J} be the atoms of the measure ω and for all ξ ∈ C0∞ (IRn ).

decompose ω = ω0 + j∈J ν (j) δx(j) , with ω0 free of atoms. Since IRn dω < ∞, J is an at most countable set. Moreover, ω0 ≥ 0. Choosing ξ such that 0 ≤ ξ ≤ 1, ξ x(j) = 1, from (4.11) we see that

46

Chapter I. The Direct Methods in the Calculus of Variations

p/q λ ≥ S ν (j) δx(j) ,

for all j ∈ J .

Since |∇k um |p − |∇k vm |p is of lower order than |∇k vm |p at points of concentration, the latter estimate also holds for μ. On the other hand, by weak lower semi-continuity we have μ ≥ |∇k u|p dx . The latter measure and the measures δx(j) being relatively singular, (2◦ ) follows.

Now, for any open set Ω ⊂ IRn such that Ω dλ ≤ S, by (4.11) with ξ = ξk ∈ C0∞ (Ω) converging to the characteristic function of Ω as k → ∞, we have p/q −1 dω ≤ dω ≤S dλ ≤ 1 . (4.12) Ω

Ω

Ω

That is, ω is absolutely continuous with respect to λ and by the RadonNikodym theorem there exists f ∈ L1 (IRn ; λ) such that dω = f dλ, λ-almost everywhere. Moreover, for λ-almost every x ∈ IRn we have dω Bρ (x) . f (x) = lim ρ→0 dλ Bρ (x) But then by (4.12), if x is not an atom of λ, ⎛ p/q ⎞ q−p q dω S B (x) ρ ⎜ ⎟ p/q = lim ⎝ dλ =0, S f (x) ≤ lim ⎠ p/q

ρ→0 ρ→0 Bρ (x) dλ Bρ (x) λ-almost everywhere. Since λ has only countably many atoms and ω0 has no atoms this implies that ω0 = 0, that is, (1◦ ). Finally, we can state the following result; see P.L. Lions [3; Theorem I.1]: 4.9 Theorem. Let k ∈ IN, p > 1, kp < n, 1q = 1p − nk . Suppose (um ) is a minimizing sequence for S in Dk,p = Dk,p (IRn ) with um Lq = 1. Then (um ) up to translation and dilation is relatively compact in Dk,p . ˜ m > 0 such that for the rescaled sequence Proof. Choose x ˜m ∈ IRn , R x−x ˜m −n/q ˜ vm (x) = Rm um ˜m R there holds (4.13)

|vm | dx =

Qm (1) = sup

x∈IRn

|vm |q dx =

q

B1 (x)

B1 (0)

1 . 2

4. The Concentration-Compactness Principle

47

Since p > 1 we may assume that vm v weakly in Lq (IRn ) and weakly in Dk,p (IRn ). Consider the families of measures μm = |∇k vm |p dx, νm = |vm |q dx and apply Lemma 4.3 to the sequence (νm ). Vanishing is ruled out by our above normalization. If we have dichotomy, let λ ∈]0, 1[ be as in Lemma 4.3.(3◦ ) and 1 2 , νm as in that for ε > 0 determine R > 0, a sequence (xm ), and measures νm lemma such that 1 2 + νm ≤ νm , 0 ≤ νm 1 2 supp(νm ) ⊂ BR (xm ), supp(νm ) ⊂ IRn \ B2R (xm ) ,

1 2 lim sup dνm − λ + dνm − (1 − λ) ≤ ε . n n m→∞ IR

IR

Choosing a sequence εm → 0 with corresponding Rm > 0 and xm , upon passing to a subsequence (νm ) if necessary, we can achieve that 1 2 supp(νm ) ⊂ BRm (xm ), supp(νm ) ⊂ IRn \ B2Rm (xm )

lim sup m→∞

and

IRn

1 dνm − λ +

IRn

2 dνm − (1 − λ) = 0 .

Moreover, in view of Lemma 4.3, we may suppose that Rm → ∞ (m → ∞). Choose ϕ ∈ C0∞ (B2 (0)) with 0 ≤ ϕ ≤ 1 and such that ϕ ≡ 1 in B1 (0). For m ∈ IN let ϕm (x) = ϕ

x−xm Rm

. Since p ≥ 1 there holds

|∇k vm |p ≥ |∇k vm |p (ϕpm + (1 − ϕm )p ). By Minkowski’s inequality we have (∇k vm )ϕm Lp (IRn ) ≥ ∇k (vm ϕm )Lp (IRn ) − C

∇l vm ∇k−l ϕm Lp (IRn ) ,

l
and similarly for (1 − ϕm ) instead of ϕm . Let Am denote the annulus Am = B2Rm (xm ) \ BRm (xm ). By Young’s inequality then for any δ > 0 we obtain ∇k vm pLp (IRn ) ≥ ∇k (vm ϕm )pLp (IRn ) + ∇k (vm (1 − ϕm ))pLp (IRn ) − βm , where the error terms βm can be estimated ∇l vm ∇k−l ϕm pLp (Am ) . βm ≤ δ∇k vm pLp (IRn ) + C(δ) l
by interpolation (as in Adams [1; Theorem 4.14]) Since |∇ ϕm | ≤ we can bound % l % l−k %|∇ vm | |∇k−l ϕm |% p ≤ C Rm ∇l vm Lp (Am ) L (Am ) (4.14) l −k − k−l ≤ C K γ∇k vm Lp (Am ) + C K Rm γ vm Lp (Am ) . k−l

l−k CRm ,

48

Chapter I. The Direct Methods in the Calculus of Variations

Here γ can be chosen arbitrarily in ]0, 1], while the constant K depends only on k and n. (Note that estimate (4.14) is invariant under dilation.) Moreover, by H¨ older’s inequality 1−1 n −k −k vm Lp (Am ) ≤ Rm L (Am ) p q vm Lq (Am ) ≤ Cvm Lq (Am ) Rm # $ 1q 1 2 ≤C dνm − dνm + dνm . IRn

IRn

IRn

Hence this term tends to 0 as m → ∞, while ∇k vm pLp (Am ) ≤ vm pDk,p remains uniformly bounded. Choosing a suitable sequence γ = γm → 0, from (4.14) we thus obtain that βm ≤ o(1), where o(1) → 0 (m → ∞). Also letting δ = δm → 0 suitably, by Sobolev’s inequality we ﬁnd vm pDk,p ≥ vm ϕm pDk,p + vm (1 − ϕm )pDk,p + o(1) ≥ S (vm ϕm pLq + vm (1 − ϕm )pLq ) + o(1) ⎡ p/q dνm + ≥S⎣ 0 ≥S

IRn \B2Rm (xm )

BRm (xm )

IRn

p/q 1

p/q 1 dνm

p/q ⎤ ⎦ + o(1) dνm

+ IRn

2 dνm

+ o(1)

≥ S λp/q + (1 − λ)p/q − o(1) , where o(1) → 0 (m → ∞). But for 0 < λ < 1 and p < q we have λp/q + (1 − λ)p/q > 1, contradicting the initial assumption that vm pDk,p = um pDk,p → S. There remains the case λ = 1, that is, case (1◦ ) of Lemma 4.3. Let xm be as in that lemma and for ε > 0 choose R = R(ε) such that dνm ≥ 1 − ε . BR (xm )

If ε < 12 our normalization condition (4.13) implies BR (xm )∩B1 (0) = ∅. Hence the conclusion of Lemma 4.3.(1◦ ) also holds with xm = 0, replacing R(ε) by 2R(ε) + 1 if necessary. Thus, if νm ν weakly, it follows that dν = 1 . IRn

By Lemma 4.8 we may assume that μm μ ≥ |∇k v|p dx +

μ(j) δx(j)

j∈J

νm ν = |v| dx + q

j∈J

ν (j) δx(j)

4. The Concentration-Compactness Principle

49

for certain points x(j) ∈ IRn , j ∈ J , and positive numbers μ(j) , ν (j) satisfying p/q ≤ μ(j) , S ν (j)

for all j ∈ J .

By Sobolev’s inequality then S + o(1) =

vm pDk,p

= IRn

dμm ≥ vpDk,p +

μ(j) + o(1)

j∈J

⎛

≥ S ⎝vLq + p/q

ν

(j) p/q

⎞ ⎠ + o(1)

j∈J

where o(1) → 1 (m → ∞). By strict concavity of the map λ → λp/q now the latter will be ⎛ ≥S (4.15)

⎝vq q L

+

⎞p/q ν

(j) ⎠

+ o(1)

j∈J

=S

p/q dν

+ o(1) = S + o(1)

IRn

and equality holds if and only if at most one of the terms vLq , ν (j) , j ∈ J , is diﬀerent from 0. Note that our normalization (4.13) assures that ν (j) ≤

1 2

for all j ∈ J .

Hence all ν (j) must vanish, vLq = 1, and vm → v strongly in Lq (IRn ). But then by Sobolev’s inequality vpDk,p ≥ S and vm Dk,p → vDk,p as m → ∞. It follows that vm → v in Dk,p (IRn ), as desired. The proof is complete. As a consequence we obtain 4.10 Corollary. For any k ∈ IN, any p > 1 such that kp < n there exists a function u ∈ Dk,p (IRn ) with uLq (IRn ) = 1 and uDk,p (IRn ) = S, where 1 1 k q = p − n and where S = S(k, p, n) is the Sobolev constant. Observe that, since Lemma 4.8 requires weak convergence um u in Dk,p (IRn ) the above proof of Theorem 4.9 cannot be extended to the case p = 1. In fact, Corollary 4.10 is false in that case, as we nhave seen that the best constant for Sobolev’s embedding D1,1 (IRn ) → L n−1 (IRn ) is attained (precisely) on characteristic functions of balls; that is, in BVloc (IRn ).

50

Chapter I. The Direct Methods in the Calculus of Variations

4.11 Notes. (1◦ ) The limiting case kp = n for Sobolev’s embedding behaves strikingly diﬀerent from the case kp < n studied above. Consider k = 1 for n simplicity. By Sobolev’s embedding W k, k → W 1,n the results for k = 1 extend to the general case. However, for k > 1 the results that follow can be slightly improved, see Brezis-Waigner [1] or Ziemer [1]. By results of Trudinger [1] and Moser [3], for any Ω ⊂⊂ IRn the space 1,n into the Orlicz space of functions u: Ω → IR such that W0 (Ω) embeds n exp |u| n−1 ∈ Lp (Ω) for any p < ∞, and there exists a limiting exponent αn > 0 such that the map n (4.16) W01,n (Ω) u → exp |u| n−1 ∈ Lp (Ω) is bounded on the unit ball B1 0; W01,n (Ω) =

u∈

W01,n (Ω)

|∇u| dx ≤ 1 n

; Ω

if and only if p ≤ αn . Quite surprisingly, and in sharp contrast to the case kp < n considered previously, Carleson-Chang [1] were able to establish that n exp αn |u| n−1 dx sup u∈B1 (0;W01,n (Ω))

Ω

is attained if Ω is a ball. Struwe [17] then showed that also for domains Ω that are close to a ball in measure the supremum is attained. Finally, Flucher [1] established the existence of an extremal function on any domain, if n = 2; see also Bandle-Flucher [1]. Moreover, at least in the case that Ω is a 2-ball (Struwe [17]), even for suﬃciently small numbers p > α2 = 4π the functional exp(p|u|2 ) dx Ep (u) = Ω

admits a relative maximizer in B1 0; H01,2 (Ω) . Thus, and since Ep (u) by the result of Moser is unbounded on B1 0; H01,2 (Ω) , we are led to expect the existence of a further critical point of saddle type, for any p > 4π suﬃciently close to 4π. In Struwe [17] such “unstable” critical points were, in fact, shown to exist for almost all such numbers p > 4π by minimax methods as we shall describe in Chapter II. Note that for p > 4π we are dealing with a “supercritical” variational problem and quite delicate techniques are needed to overcome the possible “loss of compactness”. It is an open problem whether a similar result holds for all p > 4π suﬃciently close to 4π and for any domain; moreover, the extension of this result to higher dimensions is open. (2◦) Problems with exponential nonlinearities related to the embedding (4.16) above are studied, for instance, by Adimurthi-Yadava [1]. The embedding (4.16) is also relevant for the study of nonlinearities like

5. Ekeland’s Variational Principle

51

g(u) = e2u , arising in the 2-dimensional Kazdan-Warner problem (see Moser [5], ChangYang [1]) or in the uniformization problem, that is, the Yamabe problem for surfaces; compare Section III.4.

5. Ekeland’s Variational Principle In general it is not clear that a bounded and lower semi-continuous functional E actually attains its inﬁmum. The analytic function f (x) = arctan x, for example, neither attains its inﬁmum nor its supremum on the real line. A variant due to Ekeland [1] of Dirichlet’s principle, however, permits one to construct minimizing sequences for such functionals E whose elements um each minimize a functional Em , for a sequence of functionals Em converging locally uniformly to E. 5.1 Theorem. Let M be a complete metric space with metric d, and let E: M → IR ∪ +∞ be lower semi-continuous, bounded from below, and ≡ ∞. Then for any ε, δ > 0, any u ∈ M with E(u) ≤ inf E + ε, M

there is an element v ∈ M strictly minimizing the functional ε Ev (w) ≡ E(w) + d(v, w) . δ Moreover, we have E(v) ≤ E(u), d(u, v) ≤ δ .

Fig. 5.1. Comparing E with Ev . v is a strict minimizer of Ev if and only if the absolute downward cone of slope ε/δ with vertex at v, E(v) lies entirely below the graph of E

52

Chapter I. The Direct Methods in the Calculus of Variations

Proof. Denote α = (5.1)

ε δ

and deﬁne a partial ordering on M × IR by letting

(v, β) ≤ (v , β ) ⇔ (β − β) + α d(v, v ) ≤ 0 .

This relation is easily seen to be reﬂexive, identitive, and transitive: (v, β) ≤ (v, β) , (v, β) ≤ (v , β ) ∧ (v , β ) ≤ (v, β) ⇔ v = v , β = β , (v, β) ≤ (v , β ) ∧ (v , β ) ≤ (v , β ) ⇒ (v, β) ≤ (v , β ) . Moreover, if we denote S = {(v, β) ∈ M × IR ; E(v) ≤ β} , by lower semi-continuity of E, S is closed in M × IR. To complete the proof we need a lemma. 5.2 Lemma. S contains a maximal element (v, β) with respect to the partial ordering ≤ on M × IR such that (u, E(u)) ≤ (v, β). Proof. Let (v1 , β1 ) = (u, E(u)) and deﬁne a sequence (vm , βm ) inductively as follows: Given (vm , βm ), deﬁne Sm = {(v, β) ∈ S ; (vm , βm ) ≤ (v, β)} μm = inf{β ; (v, β) ∈ Sm } ≥ inf{E(v) ; (v, β) ∈ Sm } ≥ inf E =: μ0 . M

Note that μm ≤ βm ; moreover, Sm = {(vm , βm )}, if μm = βm . Now let (vm+1 , βm+1 ) ∈ Sm be chosen such that (5.2)

βm − βm+1 ≥

1 (βm − μm ) . 2

Note that by transitivity of ≤ the sequence Sm is nested: S1 ⊃ S2 ⊃ . . . ⊃ Sm ⊃ Sm+1 ⊃ . . . . Hence also . . . ≤ μm ≤ μm+1 ≤ . . . ≤ βm+1 ≤ βm ≤ . . .. By induction, from (5.2) we obtain βm+1 − μm+1 ≤ βm+1 − μm 1 ≤ (βm − μm ) ≤ ... ≤ 2

m 1 (β1 − μ1 ) . 2

Therefore, by deﬁnition of Sm , for any m ∈ IN and any (v, β) ∈ Sm we have m 1 , |βm − β| = βm − β ≤ βm − μm ≤ C 2 (5.3) m 1 d(vm , v) ≤ α−1 (βm − β) ≤ Cα−1 . 2

5. Ekeland’s Variational Principle

53

In particular, (vm , βm ) m∈IN is a Cauchy sequence in M × IR. Thus, by com pleteness of M , (vm , βm ) converges to some limit (v, β) ∈ m∈IN Sm . By transitivity, clearly (u, E(u)) = (v1 , β1 ) ≤ (v, β). Moreover, (v, β) is maximal. ˜ for some (˜ ˜ ∈ M × IR, then also (vm , βm ) ≤ (˜ ˜ v, β) v, β) v, β) In fact, if (v, β) ≤ (˜ ˜ ˜ v , β) in (5.3) we infer for all m, and (˜ v , β) ∈ Sm for all m. Letting (v, β) = (˜ ˜ (m → ∞), whence (˜ ˜ = (v, β), as desired. that (vm , βm ) → (˜ v , β) v, β) Proof of Theorem 5.1 (Completed). Let (v, β) be maximal in S with (u, E(u)) ≤ (v, β). Comparing with (v, E(v)) ∈ S at once yields β = E(v). By deﬁnition (5.1) the statement (u, E(u)) ≤ (v, E(v)) translates into the estimate E(v) − E(u) + α d(u, v) ≤ 0 ; in particular this implies E(v) ≤ E(u) and d(u, v) ≤ α−1 (E(u) − E(v)) ≤

δ inf E + ε − inf E = δ . M ε M

Finally, if w ∈ M satisﬁes Ev (w) = E(w) + α d(v, w) ≤ E(v) = Ev (v) , by Deﬁnition 5.1 we have (v, E(v)) ≤ (w, E(w)) . Hence w = v by maximality of (v, E(v)); that is, v is a strict minimizer of Ev , as claimed. 5.3 Corollary. If V is a Banach space and E ∈ C 1 (V ) is bounded from below, there exists a minimizing sequence (vm ) for E in V such that E(vm ) → inf E, DE(vm ) → 0 V

in V ∗

as m → ∞ .

Proof. Choose a sequence (εm ) of numbers εm > 0, εm → 0 (m → ∞). For m ∈ IN choose um ∈ V such that E(um ) ≤ inf E + ε2m . V

ε2m ,

δ = εm , u = um determine an element vm = v according to For ε = Theorem 5.1, satisfying E(vm ) ≤ E(vm + w) + εm wV for all w ∈ V . Hence E(vm ) − E(vm + w) ≤ εm → 0 , wV 0 = w V →0

DE(vm )V ∗ = lim sup as claimed.

54

Chapter I. The Direct Methods in the Calculus of Variations

In Chapter II we will re-encounter the special minimizing sequences of Corollary 5.3 as “Palais-Smale sequences”. Compactness of such sequences by Corollary 5.3 turns out to be a suﬃcient condition for the existence of a minimizer for a diﬀerentiable functional E which is bounded from below on a Banach space V . Moreover, we shall see that the compactness of Palais-Smale sequences (under suitable assumptions on the topology of the level sets of E) will also guarantee the existence of critical points of saddle type. However, before turning our attention to critical points of general type we sketch another application of Ekeland’s variational principle. Existence of Minimizers for Quasi-convex Functionals Theorem 5.1 may be used to construct minimizing sequences for variational integrals enjoying better smoothness properties than can a-priori be expected. We present an example due to Marcellini-Sbordone [1]. 5.4 Example. Let Ω be a bounded domain in IRn and let f : Ω×IRN ×IRnN → IR be a Carath´eodory function satisfying the growth and coercivity conditions (5.4) |f (x, u, p)| ≤ C 1 + |u|s + |p|s for some s > 1 , C ∈ IR and (5.5)

f (x, u, p) ≥ |p|s .

Moreover, suppose f is quasi-convex in the sense of Morrey [3]; that is, for almost every x0 ∈ Ω, u0 ∈ IRN , po ∈ IRnN , and any ϕ ∈ H01,s (Ω) there holds 1 (5.6) f x0 , u0 , p0 + Dϕ(x) dx ≥ f (x0 , u0 , p0 ). Ln (Ω) Ω Note that by Jensen’s inequality condition (5.6) is weaker than requiring f to be convex in p. For u ∈ H 1,s (Ω; IRN ) now set f x, u(x), ∇u(x) dx . E(u) = E(u; Ω) = Ω

By a result of Fusco [1], conditions (5.4) and (5.6) alone already suﬃce to ensure weak lower semi-continuity of E in H 1,t (Ω; IRN ) for any t > s; on the other hand, an example by Murat and Tartar shows that this is no longer true for t = s; see for instance Marcellini-Sbordone [1; Section 2]. But if, in addition to (5.4) and (5.6), we also assume (5.5), MarcelliniSbordone [1] succeed in ﬁnding a minimizing sequence for E for given boundary data u0 which is locally bounded in H 1,t for some t > s and therefore weakly accumulates at a minimizer of E. Their proof, which we shall presently explain, is based on Ekeland’s variational principle and Giaquinta-Modica’s [1] adaption of a result of Gehring [1], Lemma 5.6 below.

5. Ekeland’s Variational Principle

55

Finally, it was shown by Acerbi-Fusco [1] that E, in fact, for any s ≥ 1 is weakly lower semi-continuous in H 1,s (Ω; IRN ) whenever f only satisﬁes (5.4), (5.6) and the condition f (x, u, p) ≥ 0. Regularity results for minimizers of (strictly) quasi-convex functionals have been obtained by Evans [1], Evans-Gariepy [1], and Giaquinta-Modica [2]. 5.5 Theorem. Under the above hypotheses (5.4)–(5.6) on f , for any u0 ∈ H 1,s (Ω; IRN ) there is a minimizing sequence (um ) for E on {u0 }+H01,s (Ω; IRN ) which is locally bounded in H 1,t for some t > s. Proof. Choose

M = {u0 } + H01,1 (Ω; IRN )

with metric d derived from the H01,1 -norm d(u, v) = |∇u − ∇v| dx . Ω

Note that by Fatou’s lemma E: M → IR ∪ +∞ is lower semi-continuous with respect to d. Let um ∈ {u0 } + H01,s (Ω; IRN ) ⊂ M be a minimizing sequence. By Theorem 5.1, if we let ε2m = E(um ) − inf{E(u) ; u ∈ M } , δm = εm , we can choose a new minimizing sequence (vm ) in M such that each vm minimizes the functional Em (w) = E(w) + εm d(vm , w) . In particular, for each Ω ⊂⊂ Ω and any w ∈ {vm } + H01,s (Ω ; IRN ) there holds (5.7) E(vm ; Ω ) ≤ E(w; Ω ) + εm |∇w − ∇vm | dx . Ω

1 Choose x0 ∈ Ω ∈ IN, 1 ≤ ν ≤ N , choose and for R < 2 dist(x 0 , ∂Ω), N ∞ ϕ = ϕν ∈ C0 B(1+ν/N)R (x0 ) ⊂ C0∞ B2R (x0 ) satisfying 0 ≤ ϕ ≤ 1, ϕ ≡ 1 on B(1+(ν−1)/N)R (x0 ), |∇ϕ| ≤ CN/R with C independent of N and R. Let 1 vm = n vm dx L (B2R \ BR (x0 )) B2R \BR (x0 )

denote the mean value of u over the annulus B2R \ BR (x0 ). Deﬁne w = wm,ν = (1 − ϕ)vm + ϕv m . Then w = vm outside B(1+ν/N)R (x0 ) and by (5.5) and (5.7) we have |∇vm |s dx ≤ E(vm ; B(1+ν/N)R (x0 )) ≤ E w; B(1+ν/N)R (x0 ) BR (x0 ) (5.8) ϕ|∇vm | dx + |∇ϕ| |vm − v m | dx . + εm B2R (x0 )

B2R \BR (x0 )

56

Chapter I. The Direct Methods in the Calculus of Variations

By Poincar´e’s inequality B2R \BR (x0 )

|vm − v m | dx ≤ cR

B2R \BR (x0 )

|∇vm | dx;

see Theorem A.10 of Appendix A. Hence, in particular, the last term in the previous inequality is bounded by CN εm B2R (x0 ) |∇vm | dx. By choice of w and condition (5.4), moreover, we may estimate 1 + |w|s + |∇w|s dx E w; B(1+ν/N)R (x0 ) ≤ C

≤C

1 + |vm |s + ϕs |vm − v m |s dx

B(1+ν/N )R (x0 )

+C

(5.9)

B(1+ν/N )R (x0 )

(1 − ϕ)s |∇vm |s + |∇ϕ|s |vm − v m |s dx

B(1+ν/N )R (x0 )

1 + |vm |s + |∇ϕ|s |vm − v m |s dx

≤C B2R (x0 )

+C B(1+ν/N )R \B(1+(ν−1)/N )R (x0 )

|∇vm |s dx .

Note that we also used H¨older’s inequality to estimate s |v m |s dx ≤ CRn(1−s) B2R (x0 )

≤C

vm dx

|vm |s dx .

B2R (x0 )

B2R (x0 )

For u ∈ L1loc (Ω), BR = BR (x0 ) ⊂ Ω let 1 — u dx = n u dx L (BR ) BR BR denote the mean value of u over BR , etc. By the Poincar´e-Sobolev inequality, interpolation, and Young’s inequality, for any δ > 0 we can bound |∇ϕ|s |vm − v m |s dx ≤ CN s R−s — |vm − v m |s dx — B2R (x0 )

B2R \BR (x0 )

≤δ—

|∇vm | dx + C(δ, N ) s

B2R (x0 )

s |∇vm | dx

—

.

B2R (x0 )

Summing the estimates (5.8) over 1 ≤ ν ≤ N , we ﬁll the annulus B2R \ BR (x0 ) on the right of (5.9), and with a uniform constant C0 we obtain |∇vm |s dx ≤ C(N ) — 1 + |vm |s + |∇vm | dx N— BR (x0 )

+ C(δ, N )

B2R (x0 ) s

— B2R (x0 )

|∇vm | dx

+ C0 (1 + δN ) — B2R

|∇vm |s dx .

5. Ekeland’s Variational Principle

57

Choosing N ≥ 2C0 + 1, δ = N −1 > 0, upon dividing by N we ﬁnd that 1 + |vm |s + |∇vm | dx |∇vm |s dx ≤ C — — BR (x0 )

B2R (x0 )

s

+C

—

|∇vm | dx

B2R (x0 )

+θ —

|∇vm |s dx

B2R (x0 )

0 < 1 independent of x0 , R, and m. (This idea with constants C and θ = 2C2C 0 +1 in a related context ﬁrst appears in Widman [1], Hildebrandt-Widman [1].) 1/s n Letting g = |∇vm |, h = 1 + |vm |s + |∇vm | , p = s · min n−s ,s > s and observing that by Sobolev’s embedding theorem |vm |s ∈ Lp , Theorem 5.5 now follows from the next lemma, due to Giaquinta-Modica [1].

5.6 Lemma. Suppose Ω is a domain in IRn and 0 ≤ g ∈ Ls (Ω), h ∈ Lp (Ω) for some p > s > 1. Assume that for any x0 ∈ Ω and 0 < R < 12 dist(x0 , ∂Ω) there holds s s s — g dx ≤ b — h dx + c — g dx + θ — g s dx BR (x0 )

B2R (x0 )

B2R (x0 )

B2R (x0 )

with uniform constants θ < 1, b, and c independent of x0 and R. Then there exists ε > 0, depending only on b, c, θ, p, s, and n, such that g ∈ Ltloc (Ω) for s < t < s + ε ≤ p, and for any x0 ∈ Ω, 0 < R < 12 dist(x, ∂Ω), and any such t there holds ⎧ 1t 1t 1s ⎫ ⎨ ⎬ — g t dx ≤C — ht dx + — g s dx , ⎩ ⎭ BR (x0 ) BR (x0 ) B2R (x0 ) with C possibly depending also on t. The proof of Lemma 5.6 goes beyond the scope of this book; a reference is Giaquinta [1; Proposition V.1.1, p. 122]. 5.7 Note. Besides various other applications, Ekeland’s variational principle has given rise to new interpretations of the mountain-pass lemma and its variants that we discuss in Section II.6; see, for instance, De Figueiredo [1] or Mawhin-Willem [1; Chapter 4.1] for an exposition. The idea of choosing special minimizing sequences to ensure convergence towards a minimizer already appears in the work of Hilbert [1] and Lebesgue [1]. In their solution of Dirichlet’s problem they use barriers and the “CourantLebesgue lemma”, that also inspired our proof of Lemma 2.8 above, to ensure the equicontinuity and hence compactness of a suitably constructed minimizing sequence for Dirichlet’s integral. (The compactness criterion for families of continuous functions on a compact domain was known from an earlier – though unsuccessful – attempt at solving Dirichlet’s problem by Arz´ela [1] in 1897.)

58

Chapter I. The Direct Methods in the Calculus of Variations

6. Duality Let V be a Banach space and suppose G: V → IR∪+∞ is lower semi-continuous and convex. Geometrically, convexity of G is equivalent to convexity of the epigraph of G epi(G) = (v, β) ∈ V × IR ; β ≥ G(v) , while lower semi-continuity is equivalent to the closedness of epi(G). By the Hahn-Banach separation theorem any closed convex set can be represented as the intersection of the closed half-spaces which contain it, bounded by support hyperplanes. Hence, for any lower semi-continuous, convex G: V → IR ∪ +∞ there exists a set LG of continuous aﬃne maps such that for any v ∈ V there holds G(v) = sup {l(v) ; l ∈ LG } , see, for instance, Ekeland-Temam [1; Proposition I.3.1]. Moreover, at any point v ∈ V where G is locally bounded there is a “support function” lv ∈ LG such that lv (v) = G(v). The set of slopes of support functions ∂G(v) = {Dl ; l ∈ LG , l(v) = G(v)} is called the subdiﬀerential of G at v.

Fig. 6.1.

6.1 Lemma. Suppose G: V → IR ∪ +∞ is lower semi-continuous and convex. If G is (Gˆ ateaux) diﬀerentiable at v, then ∂G(v) = {DG(v)}. Conversely, if G is locally bounded near v and if ∂G(v) = {v ∗ } is single-valued, then G is Gˆ ateaux-diﬀerentiable at v with Dw G(v) = w, v ∗ for all w ∈ V . (See, for instance, Ekeland-Temam [1; Proposition I.5.3] for a proof.)

6. Duality

59

6.2 The Legendre-Fenchel transform. For a function G: V → IR ∪ +∞, G ≡ +∞, not necessarily convex, the function G∗ : V ∗ → IR ∪ +∞, given by G∗ (v ∗ ) = sup {v, v ∗ − G(v) ; v ∈ V } , deﬁnes the Legendre-Fenchel transform of G. Note that G∗ is the pointwise supremum of aﬃne maps, hence G∗ is lower semi-continous and convex; moreover, the aﬃne function lv∗ , given by l∗ (v ∗ ) = v, v ∗ − G(v) , is a support function of G∗ at v ∗ if and only if (6.1)

G∗ (v ∗ ) + G(v) = v, v ∗ .

For lower semi-continuous convex functions G the situation is, in fact, symmet∗ ric in G and G∗ . To see this, introduce G∗∗ = (G∗ ) V . Note that the Legendre˜ ≤ G implies G ˜ ∗ ≥ G∗ . Hence, Fenchel transform reverses order, that is, G ˜ ˜ ∗∗ ≤ G∗∗ . applying the transform twice preserves order: G ≤ G implies G ∗∗ Moreover, for aﬃne functions l = l. Thus, for any aﬃne map l ≤ G we have

l(v) = l∗∗ (v) ≤ G∗∗ (v) = sup v, v ∗ − sup w, v ∗ − G(w) ≤ G(v) . v ∗ ∈V ∗

w∈V

(Choose w = v inside {. . .}.) It follows that G∗∗ is the largest lower semi-continuous convex function below G, and G∗∗ = G if and only if G is lower semi-continuous and convex. Our previous discussion implies: 6.3 Lemma. Suppose G: V → IR∪+∞, G ≡ +∞, is lower semi-continuous and convex, and let G∗ be its Legendre-Fenchel transform. Then (6.1) is equivalent to either one of the relations v ∈ ∂G∗ (v ∗ ) or v ∗ ∈ ∂G(v). Proof. In fact, if v ∈ ∂G∗ (v ∗ ), there is β ∈ IR such that l∗ , given by l∗ (w∗ ) = v, w∗ − β, belongs to LG∗ and satisﬁes l∗ (v ∗ ) = G∗ (v ∗ ). Since, by deﬁnition, G∗ (v ∗ ) ≥ v, v ∗ − G(v), we conclude that G(v) ≥ β. On the other hand, since G = G∗∗ ≤ (l∗ )∗ , we have

∗ ∗ ∗ G(v) ≤ sup v, w − l (w ) = β, w∗

and it follows that G(v) = β; that is, we obtain (6.1). The same reasoning shows that the assumption v ∗ ∈ ∂G(v) implies (6.1). The converse is immediate from the deﬁnitions of G∗ and G = G∗∗ .

60

Chapter I. The Direct Methods in the Calculus of Variations

In particular, if G ∈ C 1 (V ) is strictly convex, which implies that v − w, DG(v) − DG(w) > 0

if v = w ,

then DG is injective, G∗ is ﬁnite on the range of DG, and ∂G∗ DG(v) = {v} for any v ∈ V . If, in addition, DG is strongly monotone and coercive in the sense that for all v, w ∈ V there holds v − w, DG(v) − DG(w) ≥ α v − w v − w with a non-decreasing function α: [0, ∞[→ [0, ∞[ vanishing only at 0 and such that α(r) → ∞ as r → ∞, then DG: V → V ∗ is also surjective; see, for instance, Brezis [1; Corollary 2.4, p. 31]. Moreover, ∂G∗ – and hence G∗ – are locally ateaux diﬀerentiable bounded near any v ∗ ∈ V ∗ . Thus, by Lemma 6.1, G∗ is Gˆ with DG∗ (v ∗ ) = v for any v ∗ = DG(v). Finally, from the estimate v ∗ − w∗ ≥

DG∗ (v ∗ ) − DG∗ (w∗ ), v ∗ − w∗ ≥ α DG∗ (v ∗ ) − DG∗ (w∗ ) ∗ ∗ ∗ ∗ DG (v ) − DG (w )

for all v ∗ , w∗ ∈ V ∗ it follows that DG∗ is continuous; that is, G∗ ∈ C 1 (V ∗ ). We conclude that for any strictly convex function G ∈ C 1 (V ) such that DG is strongly monotone, the diﬀerential DG is a homeomorphism of V onto its dual V ∗ . (In the following we apply these results only in a ﬁnite-dimensional setting. A convenient reference in this case is Rockafellar [1; Theorem 26.5, p. 258].) Hamiltonian Systems We now apply these concepts to the solution of Hamiltonian systems. For a Hamiltonian H ∈ C 2 (IR2n ) and the standard symplectic form J on IR2n = IRn × IRn , 0 −id J = , id 0 where id is the identity map on IRn , we consider the ordinary diﬀerential equation (6.2)

x˙ = J ∇H(x) .

Note that by anti-symmetry of J we have d H x(t) = ∇H(x) · x(t) ˙ =0, dt that is, H x(t) ≡ const. along any solution of (6.2) and any “energy-surface” H = const. is invariant under the ﬂow (6.2).

6. Duality

61

6.4 Periodic solutions. One would like to understand the global structure of the set of trajectories of (6.2) and their asymptotic behavior. This is motivated of course by celestial mechanics, where questions of “stable and ramdon motion” (see Moser [4]) also seem to be of practical importance. However, with exception of the – very particular – “completely integrable” case, this program is far too complex to be dealt with as a whole. Therefore, one is interested in sub-systems of the ﬂow (6.2) such as stationary points, periodic orbits, invariant tori, or quasi-periodic solutions. While stationary points in general do not reveal too much about the system, it turns out that – C 1 -generically at least – periodic orbits of (6.2) are dense on a compact energy surface H = const.; see Pugh-Robinson [1]. Such a result seems to have already been envisioned by Poincar´e [1; Tome 1, Article 36]. For particular systems, however, such results are much harder to obtain. In fact, the question whether any given energy surface carries a periodic solution of (6.2) has only recently been answered; see Section II.9 for more details. In this section we consider the special class of convex Hamiltonians. This class includes as a simple model case the harmonic oscillator, described by the Hamiltonian 1 H(p, q) = |p|2 + |q|2 , x = (p, q) ∈ IR2 , 2 for which (6.2) possesses periodic solutions x of any given energy H(x(t)) = β > 0, all having the same period 2π. In the general case, the following result of Rabinowitz [5] and Weinstein [2], extending earlier work of Seifert [1], holds. 6.5 Theorem. Suppose H ∈ C 1 (IR2n ) is strictly convex, non-negative and coercive with H(0) = 0. Then for any α > 0 there is a periodic solution x ∈ C 1 IR ; IR2n of (6.2) with H(x(t)) = α for all t. The period T is not speciﬁed.

Remarks. Seifert and Weinstein essentially approached problem (6.2) using diﬀerential geometric methods, that is, by interpreting solutions of (6.2) as geodesics in a suitable Riemannian or Finsler metric (the so-called Jacobi metric). Rabinowitz’ proof of Theorem 6.5 revolutionized the study of Hamiltonian systems in the large as it introduced variational methods to this ﬁeld. Note that (6.2) may be interpreted as the Euler-Lagrange equation of the functional 1 1 x, J x ˙ dt (6.3) E(x) = 2 0 on the class Cα = {x ∈ C 1 (IR; IR2n ) ; x(t + 1) = x(t),

1

H(x(t)) dt = α}. 0

62

Chapter I. The Direct Methods in the Calculus of Variations

Indeed, at a critical point x ∈ Cα of E, by the Lagrange multiplier rule there exists T = 0 such that x˙ = T J ∇H(x) , and scaling time by a factor T we obtain a T -periodic solution of (6.2) on the energy surface H = α. However, the integral (6.3) is not bounded from above or below. In fact, E is a quadratic form given by an operator x → J x˙ with inﬁnitely many positive and negative eigenvalues. Owing to this complication, actually, for a long time it was considered hopeless to approach the existence problem for periodic solutions of (6.2) via the functional (6.3). Surprisingly, by methods that will be presented in Chapter II below, and by using a delicate approximation procedure, Rabinowitz was able to overcome these diﬃculties. In fact, his result is somewhat more general than stated above as it applies to compact, strictly star-shaped energy hypersurfaces. On a compact, convex energy hypersurface – as was ﬁrst observed by Clarke [2], [4] – by duality methods his original proof can be considerably simpliﬁed and the problem of ﬁnding a periodic solution of (6.2) can be recast in a way such that a solution again may be sought as a minimizer of a suitable “dual” variational problem. See also Clarke-Ekeland [1]. This is the proof we now present. Later we shall study the existence of periodic solutions of Hamiltonian systems under much more general hypotheses; see Section II.8. Moreover, we shall study the existence of multiple periodic orbits; see Section II.5. Proof of Theorem 6.5. In a ﬁrst step we reformulate the problem in a way such that duality methods can be applied. Note that by strict convexity and coerciveness of H the level surface Sα = H −1 ({α}) bounds a strictly convex neighborhood of 0 in IR2n . Thus, for any ξ in the unit sphere S 2n−1 ⊂ IR2n there exists a unique number r(ξ) > 0 such that x = r(ξ)ξ ∈ Sα . By the implicit function theorem r ∈ C 1 (S 2n−1 ). Replace H by the function

−q q |x| , if x = 0 ˜ H(x) = αr x/|x| 0, if x = 0 where q is a ﬁxed number 1 < q < 2. ˜ ∈ C 1 (IR2n ) and is homogeneous of degree q on half-rays from Note that H ˜ = α} we have S˜α = Sα the origin. Moreover, if we let S˜α = {x ∈ IR2n ; H(x) ˜ ˜ increases in and hence ∇H(x) is proportional to ∇H(x). In fact, since H direction ∇H(x) at any point x ∈ Xα , there exists a function λ > 0 such that ˜ ∇H(x) = λ(x)∇H(x) , at any x ∈ Sα . ˜ after a change of parameter thus A periodic solution x ˜ on S˜α to (6.2) for H will yield a periodic solution

6. Duality

63

x(t) = x ˜ s(t) to (6.2) on Sα for the original function H, where s solves s˙ = λ x ˜(s) . This, incidentally, also shows that whether or not a level surface H = const. carries a periodic solution of (6.2) is a question concerning the surface and the symplectic structure J – not the particular Hamiltonian H. ˜ is strictly convex. Indeed, consider any point x = ρξ ∈ IR2n \{0} Finally, H ˜ with H(x) = β > 0. Note that 1/q β 2n ˜ ˜ Sβ = {x ∈ IR ; H(x) = β} = Sα . α Thus, the hyperplane through (x, β) ∈ IR2n+1 , parallel to the hyperplane spanned by Tx S˜β ∼ = Tr(ξ)ξ Sα ⊂ IR2n × {0} ⊂ IR2n+1 together with the vector ˜ ˜ (x, x · ∇H(x)) = (x, q H(x)) = (x, qβ) ∈ IR2n+1 , ˜ which touches the graph of H ˜ precisely at is a support hyperplane for epi(H) ˜ ≥ 0 = H(0), ˜ (x, β). Similarly, since H the hyperplane IR2n × {0} is a support ˜ is strictly convex. hyperplane at (0, 0), and H ˜ Hence in the following we may assume that H = H. Let H ∗ be the Legendre-Fenchel transform of H. Note that, since H is homogeneous on rays of degree q > 1, the function H ∗ is everywhere ﬁnite. Moreover, H ∗ (0) = 0, H ∗ ≥ 0. Also note that for a function H on IR2n which is homogeneous of degree q > 1, strict convexity implies strong monotonicity of the gradient in the sense that, with a continuous function a = a(ρ) ≥ 0, vanishing only at ρ = 0, there holds q−1 |x − y| |x| + |y| |x − y|. x − y, ∇H(x) − ∇H(y) ≥ a |x| + |y| Hence, arguing as in the discussion following Lemma 6.3, we ﬁnd that H ∗ ∈ q > 2 be the conjugate exponent of q, we have C 1 (IR2n ). Finally, letting p = q−1

(6.4)

2

3 x y H(x) 2n , ; x ∈ IR − |y|p−1 |y| |y|p 3

2 x x y 2n − H ; x ∈ IR = sup , |y|p−1 |y| |y|p−1 y ; = H∗ |y|

H ∗ (y) = sup |y|p

that is, H ∗ is homogeneous on rays of degree p > 2.

64

Chapter I. The Direct Methods in the Calculus of Variations

(At this point we should remark that the components of the variable x above include both position and momentum variables. Thus, although these are certainly related, the conjugate H ∗ of H diﬀers from the usual Legendre transform of H which customarily only involves the momentum variables.) Introduce the space

Lp0 =

y ∈ Lp [0, 1]; IR2n ;

1

y dt = 0

.

0

Now, if x ∈ C 1 [0, 1]; IR2n is a 1-periodic solution of (6.2), the function y = −J x˙ ∈ Lp0 solves the system of equations y = −J x˙ , y = ∇H(x) .

(6.5) (6.6)

Equation (6.5) can be inverted (up to an integration constant x0 ∈ IRn ) by introducing the integral operator t J y dt . (Ky)(t) = K: Lp0 → H 1,p [0, 1]; IR2n , 0

By Lemma 6.3 relation (6.6) is equivalent to the relation x = ∇H ∗ (y). That is, system (6.5), (6.6) is equivalent to the system (6.5 ) (6.6 )

x = Ky + x0 x = ∇H ∗ (y)

for some x0 ∈ IR2n . The latter can be summarized in the single equation 1 ∇H ∗ (y) − Ky · η dt = 0 , (6.7) ∀η ∈ Lp0 . 0

Indeed, if y ∈

Lp0

satisﬁes (6.7), it follows that

∇H ∗ (y) − Ky = const. = x0 ∈ IRn . Hence y solves (6.5 ), (6.6 ) for some x ∈ H 1,p [0, 1]; IR2n . Transforming 1,p back to (6.5), (6.6),

→ C 0 , and therefore from (6.6) we see that y ∈ H 2n 1 is a 1-periodic solution of (6.2). Thus, (6.2) and its weak x ∈ C [0, 1]; IR “dual” form (6.7) are in fact equivalent. Now we can conclude the proof of Theorem 6.5: We recognize (6.7) as the Euler-Lagrange equation of the functional E ∗ on Lp0 , given by 1 1 E ∗ (y) = H ∗ (y) − y, Ky dt . 2 0 Note that by (6.4) the functional E ∗ is Fr´echet-diﬀerentiable and coercive on Lp0 . Moreover, E ∗ is the sum of a continuous convex and a compact quadratic

6. Duality

65

term, hence weakly lower semi-continuous. Thus, by Theorem 1.2, a minimizer 1 y ∗ ∈ Lp0 of E ∗ exists, solving (6.7). By (6.4) the quadratic term − 0 y, Ky dt ∗ in E dominates near y = 0. Since, for instance, y, Ky > 0 for y(t) = ae2πJt ∈ Lp0 with a ∈ IR2n \ {0}, we have inf E ∗ < 0, and y ∗ = 0. By the above discussion there is a constant x0 such that the function x = Ky ∗ + x0 solves (6.2). Since y ∗ = 0, also x is non-constant; hence H(x(t)) = β > 0. ˜ is homogeneous on rays. Thus a suitably rescaled multiple x But H = H ˜ of x, α 1/q x) = α, as desired. x ˜ = ( β ) x(T ·), will satisfy (6.2) with H(˜ Periodic Solutions of Nonlinear Wave-Equations As a second example we consider the problem of ﬁnding a non-constant, timeperiodic solution u = u(x, t), 0 ≤ x ≤ π, t ∈ IR, of the problem (6.8) (6.9)

Au = utt − uxx = −u|u|p−2 u(0, · ) = u(π, · ) = 0

(6.10)

u( · , t + T ) = u( · , t)

in ]0, π[×IR

for all t ∈ IR ,

where p > 2 and the period T are given. The following result again is due to Rabinowitz [6]. 6.6 Theorem. Suppose T /π ∈ Q; then there exists a non-constant T -periodic weak solution u ∈ Lp [0, π] × IR of problem (6.8)–(6.10).

Remark. For simplicity, we consider only the case T = 2π; the general case T /π ∈ Q can be handled in a similar way. The situation, however, changes completely if T is not a rational multiple of the spatial period, in this case, a rational multiple of π. Whether or not Theorem 6.7 holds true in this case is an open problem which seems to call for techniques totally diﬀerent from those we are going to describe. (See Bobenko-Kuksin [1] or P¨ oschel [1] for some recent results in this regard.) Proof. Problem (6.8)–(6.10) can be interpreted as the Euler-Lagrange equations associated with a constrained minimization problem for the functional 1 2π π |ux |2 − |ut |2 dx dt E(u) = 2 0 0 on the space 1,2 H = u ∈ Hloc [0, π] × IR ; u satisﬁes (6.9), (6.10) , endowed with the H 1,2 -norm on Ω = [0, π] × [0, 2π], subject to the constraint uLp (Ω) = 1 .

66

Chapter I. The Direct Methods in the Calculus of Variations

However, E is unbounded on this set. Moreover, the operator A = ∂t2 − ∂x2 related to the second variation of E has inﬁnitely many positive and negative eigenvalues and also possesses an inﬁnite-dimensional kernel. Therefore – as in the case of Hamiltonian systems considered above – the direct methods do not immediately apply. In order to convert (6.8)–(6.10) into a problem that we can handle, we write (6.8) as a system (6.11)

v = Au −v = u|u|p−2 = ∇G(u) ,

(6.12)

where G(u) = p1 |u|p . Since G is strictly convex, (6.12) may be inverted using the Legendre-Fenchel transform of G, 1 1 G∗ (v) = sup {uv − |u|p ; u ∈ IR} = |v|q , p q where 1 < q < 2 is the exponent conjugate to p, satisfying Lemma 6.3 then, (6.12) is equivalent to the equation

1 p

+

1 q

= 1. By

u = ∇G∗ (−v) = −v|v|q−2 .

(6.13)

In order to invert (6.11) we need to collect some facts about the wave operator A. In our exposition we basically follow Brezis-Coron-Nirenberg [1]. The representation formula (6.14)–(6.16) is due to Lovicarow´a [1]. 6.7 Estimates for the wave operator A. For T = 2π the spectrum σ(A) and kernel N of A, acting on functions in L1 (Ω) satisfying (6.9), (6.10), can be characterized as follows: σ(A) = {j 2 − k 2 ; j ∈ IN, k ∈ IN0 } ,

N=

p(t + x) − p(t − x) ; p ∈ L1loc (IR), p(s + 2π) = p(s) for almost all s,

2π

p dx = 0

.

0

The last condition appearing in the deﬁnition of N is a normalization condition.

N is closed in L1 (Ω); moreover, given f ∈ L1 (Ω) such that Ω f ϕ dx dt = 0 for all ϕ ∈ N ∩ L∞ (Ω), there exists a unique function u ∈ C(Ω), satisfying (6.9), (6.10), such that Au = f and Ω u ϕ dx dt = 0 for all ϕ ∈ N . In fact, u is given explicitly as follows: u(x, t) = ψ(x, t) + p(t + x) − p(t − x) , where ψ is constructed from a 2π-periodic extension of f to [0, π] × IR by using the fundamental solution of the wave operator; that is,

6. Duality

(6.14) with (6.15)

1 ψ(x, t) = − 2 1 c= 2

π

t+(ξ−x)

f (ξ, τ ) dτ x

π

t−(ξ−x)

π−x , π

t+ξ

f (ξ, τ ) dτ 0

dξ + c

67

dξ .

t−ξ

Note that c is constant; here, the fact that f is orthogonal to N is used. The choice of c now guarantees that u satisﬁes the boundary condition (6.9); moreover, periodicity of f implies (6.10). Finally, choosing π 1 ψ(ξ, s + ξ) − ψ(ξ, s − ξ) dξ (6.16) p(s) = 2π 0 ensures that u is L2 -orthogonal to N , as desired. Formulas (6.14)–(6.16) determine an operator K = A−1 from the weak orthogonal complement of N

⊥ 1 ∞ N = f ∈ L (Ω) ; f ϕ dx dt = 0 for all ϕ ∈ N ∩ L (Ω) Ω

into C(Ω) satisfying the condition Kf L∞ ≤ c f L1 .

(6.17)

Moreover, for f ∈ N ⊥ ∩ Lq (Ω), q > 1, we have Kf ∈ C α (Ω), with α = 1 − 1/q > 0 and Kf C α ≤ c f Lq ;

(6.18)

in particular, K is a compact, selfadjoint linear operator of N ⊥ ∩ L2 (Ω) into itself, with eigenvalues 1/(j 2 − k 2 ), j ∈ IN, k ∈ IN0 , j = k. (6.17) and (6.18) are easy consequences of (6.14)–(6.16) and H¨older’s inequality. p and let V = N ⊥ ∩ Lq (Ω), endowed with Proof of Theorem 6.6. Fix q = p−1 q the L -norm. By (6.18) the operator K: V → Lp (Ω) is compact. Deﬁne 1 E ∗ (v) = (Kv) v dx dt ; 2 Ω

clearly E ∗ ∈ C 1 (V ). Moreover, since K is compact, it follows that E ∗ is weakly lower semi-continuous. Restrict E ∗ to the unit sphere M = {v ∈ V ; vLq = 1} in Lq and consider a minimizing sequence (vm ) for E ∗ in M . We may assume that vm v ∗ weakly in Lq , whence by weak lower semi-continuity (6.19)

E ∗ (v ∗ ) ≤ lim inf E ∗ (vm ) = inf {E ∗ (v) ; v ∈ M } < 0 . m→∞

68

Chapter I. The Direct Methods in the Calculus of Variations

(To verify the last inequality recall that K possesses also negative eigenvalues.) In particular, v ∗ = 0 and v ∗ /v ∗ Lq ∈ M . But then, since E ∗ (ρv) = 2 ∗ ρ E (v) for all v, by (6.19) we must have v ∗ Lq = 1, and v ∗ ∈ M minimizes E ∗ on M . By the Lagrange multiplier rule, v ∗ satisﬁes the equation (6.20)

Kv ∗ + μv ∗ |v ∗ |q−2 ϕ dx dt = 0 ,

for all ϕ ∈ V ,

Ω

with a Lagrange parameter μ ∈ IR. Choosing ϕ = v ∗ in (6.20) we realize that μ = −2E ∗ (v ∗ ) > 0. Scaling v ∗ suitably, we obtain a non-constant function v ∈ V , satisfying (6.20) with μ = 1. But then v satisﬁes Kv + v|v|q−2 ∈ N ∩ Lp . Letting u = −v|v|q−2 ∈ Lp , thus there exists ψ ∈ N ∩ Lp such that u = Kv + ψ , u = −v|v|q−2 . But this system of equations is equivalent to (6.11), (6.12), and we conclude that u is a non-constant solution of the equation Au + u|u|p−2 = 0 , satisfying the boundary and periodicity conditions (6.9), (6.10), as desired. 6.8 Notes. (1◦ ) Actually, the solution obtained above is of class L∞ , see Brezis-Coron-Nirenberg [1; p. 672 f.]. (2◦) Theorem 6.6 remains valid for a large class of semilinear equations (6.21)

utt − xxx + g(u) = 0

involving functions g sharing the qualitative behavior of a superlinear monomial g(u) = u|u|p−2 ; see Rabinowitz [6], Brezis-Coron-Nirenberg [1]. It is not even in general necessary that g is monotone (Coron [1]) – although clearly the proof given above cannot be extended to such a case.

7. Minimization Problems Depending on Parameters

69

In the case g is smooth (C ∞ ) and strictly increasing, it was shown by Rabinowitz [6] and Brezis-Nirenberg [1] that any bounded solution u to (6.21), with boundary conditions (6.9), (6.10), is of class C ∞ . (3◦) Further aspects of problem (6.8)–(6.10) have been studied by Salvatore [1] and Tanaka [1]. See Chapter II, Remark 7.3 and Notes 10.6 for references. Other applications of duality methods to semilinear wave equations and related problems have been given by Willem [1], [2].

7. Minimization Problems Depending on Parameters Quite often in the calculus of variations a functional to be minimized may depend on parameters. In particular, this is the case if we attempt to approximate a minimization problem with constraints by a family of unconstrained variational problems involving a “penalty term”. 7.1 Penalty method. Suppose that V is a Banach space and let E and G be non-negative functionals on V . We seek to minimize E on the set M = {u ∈ V ; G(u) = 0} of admissible functions. Without any smoothness and non-degeneracy assumptions, the set M may be very “rough”, and a direct approach to the constrained minimization problem may be rather cumbersome. Thus, instead of tackling this constrained optimization problem directly, it is often more convenient to ﬁnd approximate solutions by minimizing, for > 0, the functional E (u) = E(u) + −1 G(u),

u ∈ V.

That is, we relax the constraint G(u) = 0 and admit any u ∈ V as admissible; however, as → 0, due to the “penalty term” −1 G(u), minimizers of E will come closer and closer to M and, hopefully, will converge to a solution of the constrained minimization problem. Even if M = ∅, minimizers u of E may still converge to a generalized solution of the Euler equations in some larger function space. This, however, requires controlling the penalty term G(u ). Sometimes this can be achieved via the following lemma. 7.2 Lemma. Suppose that E : V → IR, 0 < < 1, is a family of functionals with the following properties: (1◦ ) For all ∈ ] 0, 1 [ there exists u ∈ V such that E (u ) = inf V E . (2◦ ) For each u ∈ V the map → E (u) is non-increasing. Then for almost every 0 ∈ ] 0, 1 [ the map

70

Chapter I. The Direct Methods in the Calculus of Variations

→ ν = inf E V

ν 0 ,

is diﬀerentiable at 0 with diﬀerential and there holds ∂E ∂ |=0 (u0 ) ≤ |ν0 |. The same conclusion holds if, instead of (2◦ ) , we assume that for each u ∈ V the map → E (u) is non-decreasing. Proof. Since → E (u) is monotone, so is the map → ν . By Rademacher’s theorem, the latter map therefore is diﬀerentiable at almost every 0 ∈] 0, 1 [. Finally, at any such 0 , for > 0 by assumption (2◦ ) we can estimate 0 ≤ E0 (u0 ) − E (u0 ) ≤ ν0 − ν . Dividing by − 0 , and passing to the limit → 0 , we obtain the claim. In the context of the penalty method described above, assumption (2◦ ) of the lemma is always satisﬁed. Existence of a minimizer u for E , that is, assumption (1◦ ) of the lemma, can be veriﬁed under very general conditions on E and G; compare Section 1. Thus, Lemma 7.2 applies. As an illustration we ﬁrst state an abstract result in this regard. Later we will also give a concrete application. Theorem 7.3. Suppose V is reﬂexive, and let E, G: V → IR be non-negative and weakly lower semi-continuous. Also suppose that E is coercive. Then, for any > 0 the functional E = E + −1 G attains its inﬁmum ν on V . Assume that ν = o(| ln |) as → 0. Then there exists a sequence n → 0 such that −1 n G(un ) → 0 as n → ∞ for any sequence (un ) of minimizers of En . Proof. Existence of a minimizer u of E for > 0 follows from Theorem 1.2. Moreover, for any ﬁxed 0 > 0 and 0 < 1 < 0 by assumption on ν we can estimate 0 |ν |d ≥ ess inf |ν | · | ln(1 /0 )|, (7.1) o(| ln 1 |) ≥ ν1 − ν0 ≥ 1

1 <<0

whence upon dividing by | ln 1 |, and letting 1 → 0 in (7.1), we infer that lim inf |ν | = 0 . →0

Thus, by Lemma 7.2, for a suitable sequence n → 0 and any sequence (un ) of minimizers of En there holds ∂En −1 (un ) ≤ n |ν n | → 0 (n → ∞), n G(un ) = n ∂

7. Minimization Problems Depending on Parameters

71

as claimed. Similarly, if M = ∅, by the same arguments as above we can show that for a sequence n → 0 there are minimizers (un ) of En such that, as n → ∞, (un ) converges weakly to a minimizer u0 of E in M and the penalty term decays logarithmically as → 0; that is, −1 −1 n G(un ) ≤ C| ln n |

with a uniform constant C. Harmonic Maps with Singularities Let Ω ⊂ IR2 be smooth and bounded and assume (for simplicity) that Ω is simply connected. We can identify ∂Ω ∼ = S 1 via a parametrization of the boundary curve. Given a smooth map u0 : ∂Ω → S 1 ⊂ IR2 , let (Ω; IR2 ) = u ∈ H 1,2 (Ω; IR2 ); u = u0 on ∂Ω . Hu1,2 0 (Ω; IR2 ) of Dirichlet’s integral We seek to ﬁnd a minimizer u ∈ Hu1,2 0 1 |∇u|2 dx E(u) = 2 Ω subject to the constraint |u| = 1 , that is, a harmonic map u: Ω → S 1 with prescribed Dirichlet data u = u0 on ∂Ω. However, if the topological degree d of u0 , considered as a map u0 : S 1 → S 1 , is non-zero, there is no extension of u0 to a harmonic map u: Ω → S 1 of class H 1,2 . In fact, by a result of H´elein [1] and its generalization to manifolds with boundary by Qing [1] such a map would be smooth. But, if d = 0, clearly ¯ → S 1 such that u = u0 on ∂Ω. That is, any there cannot be a C 1 -map u: Ω ¯ extension of u0 to a (weakly) harmonic map must have singularities in Ω. 7.4 The Ginzburg Landau model. Bethuel-Brezis-H´elein proposed to study the resulting singular variational problem via the Ginzburg-Landau model. That is, for > 0 they study the minimizers of the functional E (u) = E(u) + −2 G(u) , where

1 G(u) = 4

u ∈ Hu1,2 (Ω; IR2 ) , 0

(1 − |u|2 )2 dx Ω

penalizes the violation of the target constraint.

72

Chapter I. The Direct Methods in the Calculus of Variations

(Choosing −2 instead of −1 gives > 0 a meaning as a “characteristic length” for the penalized problem; in our context, however, this is of no importance.) In a series of papers, summarized in their monograph Bethuel-BrezisH´elein [1], they study the properties of minimizers of the Ginzburg-Landau energy and their convergence to a harmonic map with singularities. Moreover, they derive a renormalized energy functional whose minimizers give the possible locations of the singularities of the limiting harmonic map. Three estimates, in particular, are needed for the proof of the convergence (Ω; IR2 ) be minimizers of E for > 0. We may suppose result. Let u ∈ Hu1,2 0 d > 0. Constants may depend on Ω and u0 . It is rather straightforward to establish the following upper bound (7.2)

E (u ) ≤ πd| ln | + C .

Establishing the corresponding lower bound (7.3)

E (u ) ≥ πd| ln | − C

is considerably more diﬃcult, while the ﬁnal estimate (7.4)

lim sup −2 G(u ) ≤ C →0

for the convergence proof seems to be the most delicate. Using “Poho˘zaev’s identity”, Bethuel-Brezis-H´elein proved this estimate for star-shaped domains; compare Lemma III.1.4. For general domains, the result is due to Struwe [22], [23]. For a sequence n → 0, estimate (7.4) is an immediate consequence of the “easy” upper bound (7.2) and Lemma 7.1; see Struwe [22]. 7.5 Lemma. Suppose the estimate (7.2) holds true with a uniform constant C. Then we have lim inf −2 G(u ) ≤ πd. →0

Proof. Let ν = inf Hu1,2 (Ω;IR2 ) E ≥ 0. From (7.2) and Lemma 7.2, similar to 0 (7.1) we deduce that 1 d πd| ln | + C ≥ (|ν |) 1 ∂E d (u ) ≥ ∂ 1 −2 d G(u ) =2 , and the assertion follows.

7. Minimization Problems Depending on Parameters

73

In combination with the lower bound (7.3), the same ideas actually yield (7.4) for any sequence of minimizers. This observation is due to del Pino-Felmer [1]. 7.6 Lemma. Suppose the estimates (7.2) and (7.3) hold true with a uniform constant C. Then we have lim sup −2 G(u ) ≤ C. →0

Proof. Instead of diﬀerentiating, we now take ﬁnite diﬀerences. Taking a minimizer u of E as comparison function for E2 and using (7.2), (7.3), we obtain 3 −2 G(u ) = E (u ) − E2 (u ) 4 ≤ πd| ln | + C − (πd| ln(2)| − C) ≤ πd ln 2 + C , as claimed. Thus, the proof of estimate (7.4) is complete. The arguments of Bethuel-BrezisH´elein then ﬁnally yield the following convergence result. 7.7 Theorem. For any sequence n → 0 and any sequence of minimizers un of En there exist exactly d points x1 , . . . , xd ∈ Ω such that, as n → ∞, a 1,2 ¯ weakly in Hloc (Ω \ {x1 , . . . , xd }; IR2 ) and in H 1,p (Ω, IR2 ) subsequence un u for any p < 2. The limit map u ¯: Ω \ {x1 , . . . , xd } → S 1 is harmonic and is given by u ¯(x) = eiφ(x)

d 4 x − xj |x − xj | j=1

with some harmonic function φ. (We identify IR2 ∼ = C.) The singularities x1 , . . . , xd are minima of a renormalized energy functional given explicitly in terms of u0 and the Green’s function on Ω.

7.8 Notes. The idea of using variations of parameters in order to obtain a-priori bounds on critical points (in fact, saddle points) of variational integrals depending monotonically on a real parameter was introduced by Struwe [16], in the context of surfaces of prescribed mean curvature H ∈ IR. Further applications are given in Struwe [17], [21], and Ambrosetti-Struwe[2]. Related ideas have been proposed by Schechter-Tintarev [1]. See the Preface to the Third Edition for more recent developments and references. In Section II.9 we return to the topic of parameter dependence in the context of saddle points and give applications of Hamiltonian systems.

Chapter II

Minimax Methods

In the preceding chapter we have seen that (weak sequential) lower semicontinuity and (weak sequential) compactness of the sub-level sets of a functional E on a Banach space V suﬃce to guarantee the existence of a minimizer of E. To prove the existence of saddle points we will now strengthen the regularity hypothesis on E and in general require E to be of class C 1 (V ), that is continuously Fr´echet diﬀerentiable. In this case, the notion of critical point is deﬁned and it makes sense to classify such points as relative minima or saddle points as we did in the introduction to Chapter I. Moreover, we will impose a certain compactness assumption on E, to be stated in Section 2. First, however, we recall a classical result in ﬁnite dimensions.

1. The Finite Dimensional Case In the ﬁnite dimensional case, the existence of saddle points can be obtained for instance as follows (see, for instance, Courant [1; p. 223 ﬀ.]): 1.1 Theorem. Suppose E ∈ C 1 (IRn ) is coercive and suppose that E possesses two distinct strict relative minima x1 and x2 . Then E possesses a third critical point x3 which is not a relative minimizer of E and hence distinct from x1 , x2 , characterized by the minimax principle E(x3 ) = inf max E(x) =: β , p∈P x∈p

where P = {p ⊂ IRn ; x1 , x2 ∈ p, p is compact and connected} is the class of “paths” connecting x1 and x2 . Proof. Let (pm ) be a minimizing sequence in P , satisfying max E(x) → β

x∈pm

(m → ∞) .

Since E is coercive, the sets pm are uniformly bounded. Therefore

1. The Finite Dimensional Case

p=

75

pl ,

m∈IN l≥m

the set of accumulation points of (pm ), is the intersection of a decreasing sequence of compact and connected sets, hence is compact and connected. Moreover, by construction x1 , x2 ∈ pm for every m. Hence x1,2 ∈ p, and p ∈ P . Thus max E(x) = β . max E(x) ≥ inf p ∈P x∈p

x∈p

By continuity, on the other hand, max E(x) ≤ lim sup( max E(x)) = β , x∈p

m→∞

x∈pm

and maxx∈p E(x) = β. Note that, in particular, since x1,2 are strict relative minima joined by p, it now also follows that β > max{E(x1 ), E(x2)}. To see that there is a critical point x3 ∈ p such that E(x3 ) = β, we argue indirectly: Note that by continuity of E and compactness of p the set K = {x ∈ p ; E(x) = β} is compact. Suppose DE(x) = 0 for every x ∈ K. Then there is a uniform number δ > 0 such that |DE(x)| ≥ 2δ for all x ∈ K. By continuity, there exists a neighborhood Uε = {x ∈ IRn ; ∃y ∈ K : |x − y| < ε} of K such that |DE(x)| ≥ δ in Uε . Note that this implies x1 , x2 ∈ Uε . Let η be a continuous cut-oﬀ function with support in Uε such that 0 ≤ η ≤ 1 and η ≡ 1 in a neighborhood of K. Let ∇E(x) denote the gradient of E at x, characterized by the condition ∇E(x) · v = DE(x)v for all v ∈ IRn . Deﬁne a continuous map Φ: IRn × IR → IRn by letting Φ(x, t) = x − tη(x) ∇E(x) . Note that Φ is continuously diﬀerentiable in t and d E (Φ(x, t)) = − < η(x)∇E(x), DE(x) >= −η(x)|∇E(x)|2. dt t=0 Moreover, |∇E(x)|2 ≥ δ 2 > 0 on supp(η) ⊂ Uε . By continuity then, there exists T > 0 such that d η(x) E (Φ(x, t)) ≤ − |∇E(x)|2 dt 2 for all t ∈ [0, T ], uniformly in x. Thus if we choose pT = {Φ(x, T ) ; x ∈ p} ,

76

Chapter II. Minimax Methods

Fig. 1.1. A mountain pass in the E-landscape

for any point Φ(x, T ) ∈ pT we compute that T d E (Φ(x, t)) dt E (Φ(x, T )) = E(x) + dt 0 T ≤ E(x) − η(x) |∇E(x)|2 , 2 and the latter is either ≤ E(x) < β, if x ∈ K, or ≤ β − Hence max E(x) < β .

T 2 2δ

< β, if x ∈ K.

x∈pT

But by continuity of Φ it follows that pT is compact and connected, while by choice of Uε and η also xi = Φ(xi , T ) ∈ pT , i = 1, 2. Hence pT ∈ P , contradicting the deﬁnition of β. Finally, if all critical points u of E in p with E(u) = β were relative ˜ of such points would be open in p and (by continuity of E minima, the set K ˜ = ∅. But p is and DE) also closed. Moreover, by the preceding argument K ˜ contradicting the fact that E(x1 ), E(x2 ) < β. This connected. Thus p = K, concludes the proof. 1.2 Interpretation. It is useful to think of E(x) as measuring the elevation at a point x in a landscape. Our two minima x1 , x2 then correspond to two villages at the deepest points of two valleys, separated from each other by a mountain ridge. If now we walk along a path p from x1 to x2 with the property that the maximal elevation E(x) at points x on p is minimal among all such paths we will cross the ridge at a mountain pass x3 which is a saddle point of E. Because of this geometric interpretation Theorem 1.1 is sometimes called the ﬁnite dimensional “mountain pass theorem”.

2. The Palais-Smale Condition

77

2. The Palais-Smale Condition From the experience in the preceding section we expect a functional to possess critical points of saddle type whenever the set of points with energy less than a certain value is disconnected or has a non-trivial topology. However, even in the ﬁnite dimensional setting of Theorem 1.1 and with suitable assumptions about the topology of the sub-level sets of E, saddle points in general need not exist unless a certain compactness property holds. This is illustrated by the following simple example. Example. Let E ∈ C 1 (IR2 ) be given by E(x, y) = exp(−y) − x2 . Then E0 := {(x, y) ; E(x, y) < 0} is disconnected, while there is no “mountain pass” of minimal height 0; see Figure 2.1. Note, however, that in the example above there is a sequence of paths pm , given by pm (t) = (t, m), −1 ≤ t ≤ 1, connecting the two components of E0 , such that E achieves its maximum on pm at points zm = (0, m), satisfying E(zm ) → 0, DE(zm ) → 0 as m → ∞. Moreover, the points (zm ) fail to have a ﬁnite accumulation point. It seems that this lack of compactness is responsible for the absence of saddle-type critical points.

Fig. 2.1. Searching in vain for an optimal mountain pass

As we have seen, one way of inducing the necessary compactness in the ﬁnite dimensional case is by requiring E to be coercive, which generalizes to the condition of bounded compactness in the inﬁnite dimensional case. However, as remarked earlier, in inﬁnite dimensions the requirements of bounded compactness and our regularity assumption E ∈ C 1 (V ) are incompatible. Moreover, we would like to apply minimax methods to functionals which in general are neither bounded from above nor below. Thus, any of the

78

Chapter II. Minimax Methods

former conditions on E is too restrictive. Instead, we will require the so-called Palais-Smale condition to be satisﬁed by E. Originally, in the work of Palais and Smale this assumption is stated as follows:

(C)

If S is a subset of V on which |E| is bounded but on which DE is not bounded away from zero, then there is a critical point in the closure of S.

(See Palais [1], [2], Smale [2], Palais-Smale [1].) We will replace condition (C) by a slightly stronger condition which is easier to work with. It is convenient to introduce the following concept. Definition. A sequence (um ) in V is a Palais-Smale sequence for E if |E(um )| ≤ c, uniformly in m, while DE(um ) → 0 as m → ∞. In terms of this deﬁnition our compactness condition may be phrased as follows. (P.-S.)

Any Palais-Smale sequence has a (strongly) convergent subsequence.

Condition (P.-S.) implies condition (C): From any set S as in condition (C), we may extract a Palais-Smale sequence, and, if the latter has a convergent subsequence, the limit point of this sequence will be a critical point in the closure of S. The converse, however, is not true: The functional E ≡ 0 satisﬁes (C) but in general will not satisfy (P.-S.). Note that (P.-S.) implies that any set of critical points of uniformly bounded energy is relatively compact, see Lemma 2.3.(1◦ ). In fact, if we were to strengthen condition (C) by this requirement, this new condition would be equivalent to (P.-S.). In ﬁnite dimensions, a large class of functionals satisfying (P.-S.) can be characterized as follows: 2.1 Proposition. Suppose E ∈ C 1 (IRn ) and assume the function DE + |E|: IRn → IR is coercive. Then (P.-S.) holds for E. Proof. If DE + |E| is coercive, clearly a Palais-Smale sequence will be bounded, hence will contain a convergent subsequence by the Bolzano-Weierstrass theorem. Example. Suppose E: IRn → IR is a quadratic polynomial E(x) =

n i,j=1

aij xi xj +

n i=1

b i xi + c

2. The Palais-Smale Condition

79

in x = (x1 , . . . , xn ) ∈ IRn , and that D2 E(x) = (aij )1≤i,j≤n is non-degenerate in the sense that the matrix (aij )1≤i,j≤n induces an invertible linear map IRn → IRn . Then E satisﬁes (P.-S.). We may ask whether a similar non-degeneracy condition in general will guarantee that a polynomial map satisﬁes (P.-S.). Suppose E is a polynomial of degree m in x = (x1 , . . . , xn ) ∈ IRn : a α xα , E(x) = |α|≤m αn 1 where α = (α1 , . . . , αn ), xα = xα 1 · · · xn , |α| = α1 + . . . + αn . Suppose n n 2 D E(x): IR × IR → IR is non-degenerate for any x ∈ IRn . Does E satisfy (P.S.)? (The answer seems to be unknown. In fact, this question seems related to a particular case of the Jacobi conjecture, a puzzling problem in algebraic geometry; see, for instance, Bass-Connell-Wright [1].) In general, we can say the following:

2.2 Proposition. Suppose that E has the following properties. (1◦) Any Palais-Smale sequence for E is bounded in V . (2◦) For any u ∈ V we can decompose DE(u) = L + K(u) , where L: V → V ∗ is a ﬁxed boundedly invertible linear map and the operator K maps bounded sets in V to relatively compact sets in V ∗ . Then E satisﬁes (P.-S.). Proof. Any (P.-S.)-sequence (um ) is bounded by assumption. Moreover DE(um ) = Lum + K(um ) → 0 implies that

um = o(1) − L−1 K(um ) ,

where o(1) → 0 in V as m → ∞. By boundedness of (um ) and compactness of K the sequences (L−1 K(um )) – and hence (um ) – are relatively compact. The Palais-Smale condition permits one to distinguish a certain family of neighborhoods of critical points of a functional E and thus oﬀers a useful characterization of regular values of E. For β ∈ IR, δ > 0, ρ > 0 let Eβ = {u ∈ V ; E(u) < β}, Kβ = {u ∈ V ; E(u) = β, DE(u) = 0} , Nβ,δ = {u ∈ V ; |E(u) − β| < δ, DE(u) < δ} , Uβ,ρ = {v ∈ V ; v − u < ρ} . u∈Kβ

80

Chapter II. Minimax Methods

That is, Kβ is the set of critical points of E having “energy” β, {Uβ,ρ }ρ>0 is the family of norm-neighborhoods of Kβ . 2.3 Lemma. Suppose E satisﬁes (P.-S.). Then for any β ∈ IR the following holds: (1◦) Kβ is compact. (2◦) The family {Uβ,δ }ρ>0 is a fundamental system of neighborhoods of Kβ . (3◦) The family {Nβ,δ }δ>0 is a fundamental system of neighborhoods of Kβ . Proof. (1◦ ) Any sequence (um ) in Kβ by (P.-S.) has a convergent subsequence. By continuity of E and DE any accumulation point of such a sequence also lies in Kβ , and Kβ is compact. (2◦) Any Uβ,ρ , ρ > 0, is a neighborhood of Kβ . Conversely, let N be any open neighborhood of Kβ . Suppose by contradiction that for ρm → 0 there is a sequence of points um ∈ Uβ,ρm \N . Let vm ∈ Kβ be such that um −vm ≤ ρm . Since Kβ is compact by (1◦ ), we may assume that vm → v ∈ Kβ . But then also um → v, and um ∈ N for large m, contrary to assumption. (3◦) Similarly, each Nβ,δ , for δ > 0, is a neighborhood of Kβ . Conversely, suppose that for some neighborhood N of Kβ and δm → 0 there is a sequence um ∈ Nβ,δm \ N . By (P.-S.) the sequence (um ) accumulates at a critical point u ∈ Kβ ⊂ N . The contradiction proves the lemma. 2.4 Remarks. (1◦ ) In particular, if Kβ = ∅ for some β ∈ IR there exists δ > 0 such that Nβ,δ = ∅; that is, the diﬀerential DE(u) is uniformly bounded in norm away from 0 for all u ∈ V with E(u) close to β. (2◦) The conclusion of Lemma 2.3 remains valid at the level β under the weaker assumption that (P.-S.)-sequences (um ) for E such that E(um ) → β are relatively compact. This observation will be useful when dealing with limiting cases for (P.-S.) in Chapter III. 2.5 Cerami’s variant of (P.-S.). Cerami [1], [2; Teorema (*), p. 166] has proposed the following variant of (P.-S.): (2.1)

Any sequence (u m ) such that |E(um )| ≤ c uniformly and DE(um ) 1 + um → 0 (m → ∞) has a (strongly) convergent subsequence.

Condition (2.1) is slightly weaker than (P.-S.) while the most important implications of (P.-S.) are retained; see Cerami [1],[2] or Bartolo-Benci-Fortunato [1]. However, for most purposes it suﬃces to use the standard (P.-S.) condition. Therefore and in order to achieve a coherent and simple exposition consistent with the bulk of the literature in the ﬁeld, in the following our presentation will be based on (P.-S.) rather than (2.1). Variants of the Palais-Smale condition for non-diﬀerentiable functionals will be discussed in a later section. (See Section 10.)

3. A General Deformation Lemma

81

3. A General Deformation Lemma Besides the compactness condition, the second main ingredient in the proof of Theorem 1.1 is the (local) gradient-line deformation Φ. Following Palais [4], we will now construct a similar deformation for a general C 1 -functional in a Banach space. The construction may be carried out in the more general setting of C 1 functionals on complete, regular C 1,1 -Banach manifolds with Finsler structures; see Palais [4]. However, for most of our purposes it suﬃces to consider a Banach space as the ambient space on which a functional is deﬁned. Pseudo-gradient Flows on Banach Spaces In the following we will initially assume that E is a C 1 -functional on a Banach space V . Moreover, we denote V˜ = {u ∈ V ; DE(u) = 0} the set of regular points of E. As a substitute for the notion of gradient (which requires an inner product to be deﬁned) we introduce the following concept. 3.1 Definition. A pseudo-gradient vector ﬁeld for E is a locally Lipschitz continuous vector ﬁeld v: V˜ → V such that the conditions (1◦) v(u) < 2 min{DE(u), 1} , (2◦) v(u), DE(u) > min {DE(u), 1} DE(u) hold for all u ∈ V˜ . Note that we require v to be locally Lipschitz. Hence, even in a Hilbert space the following result is somewhat remarkable. 3.2 Lemma. Any functional E ∈ C 1 (V ) admits a pseudo-gradient vector ﬁeld v: V˜ → V. Proof. For u ∈ V˜ choose w = w(u) such that (3.1)

w < 2 min{DE(u), 1} ,

(3.2)

w, DE(u) > min{DE(u), 1}DE(u) .

By continuity, for any u ∈ V˜ there is a neighborhood W (u) such that (3.1) and (3.2) hold (with w = w(u)) for all u ∈ W (u). Since V˜ ⊂ V is metrizable and hence paracompact, there exists a locally ﬁnite reﬁnement {Wι }ι∈I of the cover {W (u)}u∈V˜ of V˜ , consisting of neighborhoods Wι ⊂ W (uι ); see for instance Kelley [1; Corollary 5.35, p. 160]. Choose a Lipschitz continuous partition of unity {ϕι }ι∈I subordinate to {Wι }ι∈I ; that is, choose Lipschitz continuous functions 0 ≤ ϕι ≤ 1 with support in Wι and such that ι∈I ϕι ≡ 1 on V˜ . For instance, following Palais [4; p. 205 f.], we may let

82

Chapter II. Minimax Methods

ρι (u) = dist(u, V \ Wι ) = inf u − v ; v ∈ Wι and deﬁne

ρι (u) . ι ∈I ρι (u) Clearly, 0 ≤ ϕι ≤ 1, ϕι = 0 outside Wι , and ι∈I ϕι ≡ 1. Moreover, since {Wι }ι∈I is locally ﬁnite, for any u ∈ V˜ there exists a neighborhood W of u such that W ∩ Wι = ∅ for at most ﬁnitely many indices ι ∈ I, and Lipschitz continuity of ϕι on W is immediate from Lipschitz continuity of the family {ρι }ι∈I . Finally, we may let ϕι (u) w(uι ) . v(u) = ϕι (u) =

ι∈I

The relations (3.1) and (3.2) being satisﬁed by w(uι ) on the support of ϕι , their convex linear combination v is a pseudo-gradient vector ﬁeld for E, as required. 3.3 Remark. If E admits some compact group action G as symmetries, v may be constructed to be G-equivariant. In particular, if E is even, that is, E(u) = E(−u), with symmetry group {id, −id} ∼ = ZZ2 , we may choose v˜(u) =

1 (v(u) − v(−u)) , 2

where v is any pseudo-gradient vector ﬁeld for E, to obtain a ZZ2 -equivariant pseudo-gradient vector ﬁeld v for E, satisfying v(−u) = −v(u). More generally, suppose G is a compact Lie group acting on V ; that is, suppose there is a group homomorphism of G onto a subgroup – indiscriminately denoted by G – of the group of linear isometries of V such that the evaluation map G × V → V ; (g, u) → gu is continuous. Also suppose that G leaves E invariant ∀(g, u) ∈ G × V .

E(gu) = E(u),

Then it suﬃces to let v˜(u) =

g −1 v(gu) dg

G

be the average of any pseudo-gradient vector ﬁeld v for E with respect to an invariant Haar’s measure dg on G in order to obtain a G-equivariant pseudogradient vector ﬁeld v˜, satisfying v˜(gu) = g˜ v(u) for all g and u. We are now ready to state the main theorem in this section.

3. A General Deformation Lemma

83

3.4 Theorem (Deformation Lemma). Suppose E ∈ C 1 (V ) satisﬁes (P.-S.). Let β ∈ IR, ε > 0 be given and let N be any neighborhood of Kβ . Then there exist a number ε ∈]0, ε[ and a continuous 1-parameter family of homeomorphisms Φ(·, t) of V , 0 ≤ t < ∞, with the properties (1◦) Φ(u, t) = u, if t = 0, or DE(u) = 0, or |E(u) − β| ≥ ε; (2◦) E (Φ(u, t)) is non-increasing in t for any u ∈ V ; (3◦) Φ (Eβ+ε \ N, 1) ⊂ Eβ−ε , and Φ (Eβ+ε , 1) ⊂ Eβ−ε ∪ N . Moreover, Φ: V × [0, ∞[→ V has the semi-group property; that is, Φ(·, t) ◦ Φ(·, s) = Φ(·, s + t) for all s, t ≥ 0. Proof. Lemma 2.3 permits one to choose numbers δ, ρ > 0 such that N ⊃ Uβ,2ρ ⊃ Uβ,ρ ⊃ Nβ,δ . We may suppose δ, ρ ≤ 1. Let η be a locally Lipschitz continuous function on V such that 0 ≤ η ≤ 1, η ≡ 1 outside Nβ,δ , η ≡ 0 in Nβ,δ/2 . Also let ϕ be a Lipschitz continuous function on IR such that 0 ≤ ϕ ≤ 1, ϕ(s) ≡ 0, if |β − s| ≥ min{ε, δ/4}, ϕ(s) ≡ 1, if |β − s| ≤ min{ε/2, δ/8}. Finally, let v: V˜ → V be a pseudo-gradient vector ﬁeld for E. Deﬁne

˜ e(u) = −η(u) ϕ (E(u)) v(u), if u ∈ V . 0, else By choice of ϕ and η, the vector ﬁeld e vanishes identically (and therefore is Lipschitz continuous) near critical points u of E. Hence e is locally Lipschitz continuous throughout V . Moreover, since v < 2 uniformly, also e ≤ 2 is uniformly bounded. Hence there exists a global solution Φ: V × IR → V of the initial value problem ∂ Φ(u, t) = e (Φ(u, t)) ∂t Φ(u, 0) = u . Φ is continuous in u, diﬀerentiable in t and has the semi-group property Φ(·, s)◦ Φ(·, t) = Φ(·, s + t), for any s, t ∈ IR. In particular, for any t ∈ IR the map Φ(·, t) is a homeomorphism of V . Properties (1◦ ) and (2◦ ) are trivially satisﬁed by construction and the properties of v. Moreover, for ε ≤ min{ε/2, δ/8} and u ∈ Eβ+ε , if E(Φ(u, 1)) ≥ β−ε it follows from (2◦ ) that |E (Φ(u, t)) − β| ≤ ε and hence that ϕ E Φ(u, t) = 1 for all t ∈ [0, 1]. Diﬀerentiating, by the chain rule we thus obtain 1 d E (Φ(u, t)) dt E Φ(u, 1) = E(u) + 0 dt 1 <β+ε− η(Φ(u, t))v(Φ(u, t)), DE(Φ(u, t)) dt (3.3) 0 v(Φ(u, t)), DE(Φ(u, t)) dt ≤β+ε− {t;Φ(u,t) ∈Nβ,δ }

≤ β + ε − L1 {t; Φ(u, t) ∈ Nβ,δ } · δ 2 .

84

Chapter II. Minimax Methods

But if either u ∈ N or Φ(u, 1) ∈ N , by uniform boundedness e ≤ 2 and since V \N and Nβ,δ are separated by the “annulus” Uβ,2ρ \Uβ,ρ of width ρ, certainly ρ L1 {t ; Φ(u, t) ∈ Nβ,δ } ≥ . 2

(3.4) Hence, if we choose ε ≤

δ2 ρ 4 ,

estimate (3.3) gives

E (Φ(u, 1)) < β + ε −

ρδ 2 ≤β−ε , 2

and (3◦ ) follows. 3.5 Remarks. (1◦ ) Since the deformation Φ: V × [0, ∞[→ V is obtained by integrating a suitably truncated pseudo-gradient vector ﬁeld, Φ will be called a (local) pseudo-gradient ﬂow. (2◦) If Kβ = ∅, we may choose N = ∅ and hence obtain a uniform reduction of energy near β in this case. (3◦) The Palais-Smale condition was only used to obtain estimate (3.4). For this it is enough to assume that (P.-S.) holds at the level β, see Remark 2.4(2◦ ). In particular, if N = Kβ = ∅ the conclusion of Theorem 3.4 remains valid if condition (P.-S.)is replaced by the assumption that Nβ,δ = ∅ for some δ > 0. (4◦) If E is invariant under a compact group action G, as in Remark 3.3, we can achieve that Φ is G-equivariant in the sense that there holds Φ(gu, t) = g Φ(u, t)

for all u ∈ V, g ∈ G, t ≥ 0 .

3.6 Comparison with gradient flows. It may be of interest to consider the special case of a C 2 -functional E on a real Hilbert space H with scalar product (·, ·) and induced norm · . In this case, a gradient vector ﬁeld ∇E: H → H is deﬁned as in the ﬁnite dimensional caseby letting ∇E(u) at any u ∈ H be the unique vector in H such that ∇E(u), v = DE(u)v for all v ∈ H, equivalently characterized by (3.5)

∇E(u) = DE(u), ∇E(u), DE(u) = DE(u)2 .

Moreover, since ∇E(u) − ∇E(v) = DE(u) − DE(v) , if E ∈ C 2 , ∇E is of class C 1 and deﬁnes a local gradient ﬂow Φ by letting ∂ Φ(u, t) = −∇E (Φ(u, t)) ∂t Φ(u, 0) = u . To interpret Φ we identify E with its graph G(E) = {(u, E(u)) ∈ H ×IR}. Then in the picture outlined in Interpretation 1.2 the ﬂow-lines of Φ become paths

3. A General Deformation Lemma

85

of steepest descent, and the rest points of Φ are precisely the critical points of E where G(E) has a horizontal tangent plane. In a Banach space V as ambient space, note that in general by (3.5) a gradient vector need not be uniquely determined unless V is uniformly locally convex. Moreover, also in this case, the duality map j: V ∗ → V , which maps v ∈ V ∗ to w = j(v) ∈ V satisfying w, v = v2 = w2 , in general is only uniformly continuous on bounded sets but may fail to be Lipschitz. Fortunately, it is not at all necessary to deform along lines of “steepest” descent to obtain existence results for saddle points, “steep enough” suﬃces. The notions of pseudo-gradient vector ﬁeld and pseudo-gradient ﬂow allow for the necessary ﬂexibility. Pseudo-Gradient Flows on Manifolds The above constructions of pseudo-gradient vector ﬁelds and pseudo-gradient ﬂows can easily be generalized to the setting of a C 1 -functional on a complete C 1,1 -Finsler manifold. We basically follow Palais [4]. 3.7 Finsler manifolds. Let F be a Banach space bundle over a space M and let · be a continuous real valued function on F such that the restriction · u of · to each ﬁber Fu is an admissible norm for Fu . If we trivialize F in a neighborhood of a point u0 ∈ M , using Fu0 as the standard ﬁber, then for each u near u0 the norm · u becomes a norm on Fu0 . We say · is a Finsler structure for the bundle F if for any ε > 0, each u0 ∈ M , and each such trivialization in an atlas deﬁning the bundle structure of E vu , v uo v∈Fu0 \{0} sup

vu0 <1+ε , v∈F \{0} vu sup

if u is suﬃciently near u0 . Recall that a topological space M is regular if for each point x ∈ M and ˜ of u such that any neighborhood U of u there is a closed neighborhood U r ˜ U ⊂ U . Now, a Finsler manifold of class C , r ≥ 1, is a regular C r -Banach manifold M , modeled on a Banach space V , together with a Finsler structure · on the tangent bundle T M . Then also the co-tangent space T ∗ M carries a natural Finsler structure, indiscriminately denoted by · , characterized by letting v ∗ u = sup |v, v ∗ | ; v ∈ Tu M, vu ≤ 1 for any v ∗ ∈ Tu∗ M . Finally, · induces a metric % 1% % %d % % (3.5) d(u, v) = inf dt , % dt p(t)% p 0 p(t) where the inﬁmum is taken over all C 1 -paths p: [0, 1] → M joining p(0) = u, p(1) = v; see Palais [4; p. 208 ﬀ.]. We say that M is complete if M is complete with respect to this metric d.

86

Chapter II. Minimax Methods

As an example of a complete C r -Finsler manifold we may consider any (norm-) complete C r -submanifold M of a Banach space V , with Tu M carrying the norm induced by the inclusion Tu M ⊂ Tu V ∼ =V. 3.8 Definition. Let M be a C 1,1 Finsler manifold modeled on a Banach space ˜ = {u ∈ M ; DE(u) = 0} the set of regular points of E. A V , E ∈ C 1 (M ), M pseudo-gradient vector ﬁeld for E is a Lipschitz continuous vector ﬁeld (section) ˜ → T M in the tangent bundle T M with the property that v(u) ∈ Tu M and v: M (1◦) v(u)u < 2 min{DE(u)u , 1}, (2◦) v(u), DE(u) > min{DE(u)u , 1}DE(u)u, ˜ , where · u denotes the norm in the tangent space Tu M ∼ for all u ∈ M = V to M at the point u. ˜ there exists a (constant) pseudoAs in the proof of Lemma 3.2, for any u ∈ M gradient vector ﬁeld in a local trivialization of the tangent bundle around u. Since a C 1,1 -Finsler manifold is paracompact, see Palais [4; p. 203], and a paracompact C 1,1 -Banach manifold always admits locally Lipschitz partitions of unity, see Palais [4; p. 205 f.], a family of local pseudo-gradient vector ﬁelds as above may be patched together to yield a pseudo-gradient vector ﬁeld ˜ → T M just as in the “ﬂat” case M = V . Thus we obtain (Palais [4; 3.3. v: M p. 206]): 3.9 Lemma. Any functional E ∈ C 1 (M ) on a C 1,1 -Finsler manifold M admits ˜ → TM. a pseudo-gradient vector ﬁeld v: M 3.10 Remark. Again, if G is a compact Lie group acting (smoothly) through isometries on M and if E is G-invariant, v may be constructed to be Gequivariant in the sense that v g(u) = Dg(u) v(u) , ∀u ∈ M, g ∈ G . Indeed, we may let

−1

(Dg(u))

v˜(u) =

v (g(u)) dg

G

with respect to an invariant measure dg on G. (Note that in the “ﬂat” case M = V , with G acting through linear isometries, the tangent map Dg(u): TuM ∼ = V → Tg(u) M ∼ = V may be identiﬁed with the map g itself.) For β ∈ IR, δ, ρ > 0 deﬁne Kβ , Nβ,δ , Uβ,δ as in the ﬂat case, using the Finsler structure to deﬁne the norm DE(u)u and the distance (3.5). The Palais-Smale condition can now be stated as in Section 2, and Lemma 2.3 remains true for E ∈ C 1 (M ) on a C 1,1 -Finsler manifold M . Then the construction in the proof of Theorem 3.4 can be carried over to obtain a local pseudo-gradient ﬂow Φ: D(Φ) ⊂ M × [0, ∞[→ M for E. Note that for Φ to be deﬁned globally on M × [0, ∞[ we also need to assume that M is complete with respect to the metric (3.5). This yields the following result.

4. The Minimax Principle

87

3.11 Theorem. Suppose M is a complete C 1,1 -Finsler manifold and E ∈ C 1 (M ) satisﬁes (P.-S.). Let β ∈ IR, ε > 0 be given and let N be any neighborhood of Kβ . Then there exist a number ε ∈]0, ε[ and a continuous 1-parameter family of homeomorphisms Φ(·, t) of M, 0 ≤ t < ∞, with the properties (1◦) Φ(u, t) = u, if t = 0, or DE(u) = 0, or |E(u) − β| ≥ ε; (2◦) E (Φ(u, t)) is non-increasing in t for any u ∈ M ; (3◦) Φ(Eβ+ε \ N, 1) ⊂ Eβ−ε , and Φ(Eβ+ε , 1) ⊂ Eβ−ε ∪ N . Moreover, Φ has the semi-group property Φ(·, s) ◦ Φ(·, t) = Φ(·, s + t), ∀s, t ≥ 0. If M admits a compact group of symmetries G and if E is G-invariant, Φ can be constructed to be G-equivariant, that is, such that Φ(g(u), t) = g (Φ(u, t)) for all g ∈ G, u ∈ M, t ≥ 0. 3.12 Remarks. (1◦ ) It suﬃces to assume that (P.-S.) is satisﬁed at the level β, see Remark 2.4(2◦ ). In particular, if N = Kβ = ∅, condition (P.-S.) may be replaced by the assumption that Nβ,δ = ∅ for some δ > 0, see Remark 3.5(3◦ ). (2◦ ) Completeness of M is only needed to ensure that the trajectories of the pseudo-gradient ﬂow Φ are complete in forward direction.

4. The Minimax Principle The “deformation lemma” Theorem 3.4 is a powerful tool for proving the existence of saddle points of functionals under suitable hypotheses on the topology of the manifold M or the sub-level sets Eα of E. We proceed to state a very general result in this direction: the generalized minimax principle of Palais [4; p. 210]. Later in this section we sketch some applications of this result. 4.1 Definition. Let Φ: M × [0, ∞[→ M be a semi-ﬂow on a manifold M . A family F of subsets of M is called (positively) Φ-invariant if Φ(F, t) ∈ F for all F ∈ F , t ≥ 0. 4.2 Theorem. Suppose M is a complete Finsler manifold of class C 1,1 and E ∈ C 1 (M ) satisﬁes (P.-S.). Also suppose F ⊂ P(M ) is a collection of sets which is invariant with respect to any continuous semi-ﬂow Φ: M × [0, ∞[→ M such that Φ(·, 0) = id, Φ(·, t) is a homeomorphism of M for any t ≥ 0, and E (Φ(u, t)) is non-increasing in t for any u ∈ M . Then, if β = inf sup E(u) F ∈F u∈F

is ﬁnite, β is a critical value of E. Proof. Assume by contradiction that β ∈ IR is a regular value of E. Choose ε = 1, N = ∅ and let ε > 0, Φ: M × [0, ∞[→ M be determined according to Theorem 3.11. By deﬁnition of β there exists F ∈ F such that

88

Chapter II. Minimax Methods

sup E(u) < β + ε ; u∈F

that is, F ⊂ Eβ+ε . By property 3.11(3◦ ) of Φ and invariance of F , if we let F1 = Φ(F, 1), we have F1 ∈ F and F1 ⊂ Eβ−ε ; that is, sup E(u) ≤ β − ε , u∈F1

which contradicts the deﬁnition of β. Of course, it would be suﬃcient to assume that condition (P.-S.) is satisﬁed at the level β and that F is forwardly invariant only with respect to the pseudogradient ﬂow. In the above form, the minimax principle can be most easily applied if E is a functional on a manifold M with a rich topology. But also in the “ﬂat” case E ∈ C 1 (V ), V a Banach space, such topological structure may be hidden in the sub-level sets Eγ of E. 4.3 Examples. (Palais, [3; p. 190 f.]) Suppose M is a complete Finsler manifold of class C 1,1 , and E ∈ C 1 (M ) satisﬁes (P.-S.). (1◦) Let F = {M }. Then F is invariant under any semi-ﬂow Φ. Hence, if β = inf sup E(u) = sup E(u) F ∈F u∈F

u∈M

is ﬁnite, β = maxu∈M E(u) is attained at a critical point of E. (2◦) Let F = {{u}; u ∈ M }. Then, if β = inf sup E(u) = inf E(u) F ∈F u∈F

u∈M

is ﬁnite, β = minu∈M E(u) is attained at a critical point of E. (3◦) Let X be any topological space, and let [X, M ] denote the set of free homotopy classes [f ] of continuous maps f : X → M . For given [f ] ∈ [X, M ] let F = {g(X) ; g ∈ [f ]} . Since [Φ ◦ f ] = [f ] for any homeomorphism Φ of M homotopic to the identity, the family F is invariant under such mappings Φ. Hence, if β = inf sup E(u) F ∈F u∈F

is ﬁnite, β is a critical value. (4◦) Let Hk (M ) denote the k-dimensional homology of M (with arbitrary coeﬃcients). Given a non-trivial element f ∈ Hk (M ), denote by F the collection of all F ⊂ M such that f is in the image of Hk (iF ): Hk (F ) → Hk (M ) ,

4. The Minimax Principle

89

where Hk (iF ) is the homomorphism induced by the inclusion iF : F → M . Then F is invariant under any homeomorphism Φ homotopic to the identity, and by Theorem 4.2, if β = inf sup E(u) F ∈F u∈F

is ﬁnite, then β is a critical value. There is a “dual version” of (4◦ ): (5◦) If H k is any k-dimensional cohomology functor, f a non-trivial element f ∈ H k (M ),

f = 0 ,

let F denote the family of subsets F ⊂ M such that f is not annihilated by the restriction map H k (iF ): H k (M ) → H k (F ) . Then, if β = inf sup E(u) F ∈F u∈F

is ﬁnite, β is a critical value. In the more restricted setting of a functional E ∈ C 1 (V ), similar results are valid if M is replaced by any sub-level set Eγ , γ ∈ IR. We leave it to the reader to ﬁnd the analogous variants of (3◦ ), (4◦ ), and (5◦ ). See also Ghoussoub [1]. Closed Geodesics on Spheres A diﬀerent construction of a ﬂow-invariant family is at the basis of the next famous result. We assume the notion of geodesic to be familiar from diﬀerential geometry. Otherwise the reader may regard (4.2) below as a deﬁnition. 4.4 Theorem. (Birkhoﬀ [1]) On any compact surface S in IR3 which is C 3 -diﬀeomorphic to the standard sphere, there exists a non-constant closed geodesic. d u. Deﬁne Proof. Denote u˙ = dt 1,2 IR; IR3 ; u(t) = u(t + 2π), H 1,2 IR/2π; S = u ∈ Hloc u(t) ∈ S for almost every t ∈ IR

the space of closed curves u: IR/2π → S with ﬁnite energy 1 2π 2 |u| ˙ dt . E(u) = 2 0 By H¨older’s inequality (4.1) |u(s) − u(t)| ≤

t

|u| ˙ dτ ≤ s

1/2 t 1/2 |t − s| |u| ˙ 2 dτ ≤ (2|t − s|E(u)) , s

90

Chapter II. Minimax Methods

functions u ∈ H 1,2 (IR/2π; S) with E(u) ≤ γ will be uniformly √ H¨ older continuous with H¨ older exponent 1/2 and H¨ older norm bounded by 2πγ. Hence, if S ∈ C 3 , H 1,2 (IR/2π; S) becomes a complete C 2 -submanifold of the Hilbert space H 1,2 (IR/2π; IR3 ) with tangent space Tu H 1,2 (IR/2π; S) = {ϕ ∈ H 1,2 (IR/2π; IR3 ) ; ϕ(t) ∈ Tu(t) S ∼ = IR2 } at any closed curve u ∈ H 1,2 (IR/2π; S); see, for instance, Klingenberg [1; Theorem 1.2.9]. By (4.1), if E(u) is suﬃciently small so that the image of u is covered by a single coordinate chart, of course Tu H 1,2 (IR/2π; S) ∼ = H 1,2 (IR/2π; IR2 ) . Moreover, E is analytic on H 1,2 (IR/2π; IR3 ), hence as smooth as H 1,2 (IR/2π; S) when restricted to that space. At a critical point u ∈ C 2 , upon integrating by parts, 2π 2π u˙ ϕ˙ dt = −¨ u ϕ dt = 0 , ∀ϕ ∈ Tu H 1,2 (IR/2π; S) ; (4.2) 0

0

that is, u ¨(t) ⊥ Tu(t) S for all t, which is equivalent to the assertion that u is a geodesic, parametrized by arc length. More generally, at a critical point u ∈ H 1,2 (IR/2π; S), if n: S → IR3 denotes a (C 2 -) unit normal vector ﬁeld on S, for any ϕ ∈ H 1,2 (IR/2π; IR3 ) we have ϕ − n(u) n(u) · ϕ ∈ Tu H 1,2 (IR/2π; S) , where · is the inner product in IR3 . Inserting this into (4.2) and observing that · (4.3) u ¨ · n(u) = u˙ · n(u) − u˙ · Dn(u) u˙ = −u˙ · Dn(u) u˙ in the distribution sense, we obtain that (4.4) u ¨ + u˙ · Dn(u) u˙ n(u) = 0 . From (4.4) we now obtain that u ∈ H 2,1 (IR/2π) → C 1 (IR/2π). Hence, by (4.4) again, u ∈ C 2 (IR/2π), and our previous discussion shows that closed geodesics on S exactly corresponding to the critical points of E on H 1,2 (IR/2π; S). Moreover, E satisﬁes the Palais-Smale condition on H 1,2 (IR/2π; S): If (um ) is a sequence in H 1,2 (IR/2π; S) such that E(um ) ≤ c < ∞ and 2π →0, sup u ˙ ϕ ˙ dt DE(um ) = m ϕ∈Tum H 1,2 (IR/2π;S) ϕ1,2 ≤1

0

then (um ) contains a strongly convergent subsequence.

4. The Minimax Principle

91

Proof of (P.-S.). Since E(um ) ≤ c uniformly, by (4.1) the sequence (um ) is equi-continuous. But S is compact, in particular bounded; hence (um ) is equibounded. Thus, by Arz´ela-Ascoli’s theorem we may assume that um → u uniformly and weakly in H 1,2 (IR/2π; IR3 ). It follows that u ∈ H 1,2 (IR/2π; S). Via the unit normal vector ﬁeld n we can deﬁne the projection πv : H 1,2 (IR/2π; IR3 ) → Tv H 1,2 (IR/2π; S) by letting ϕ → πv ϕ(t) = ϕ(t) − n (v(t)) n (v(t)) · ϕ(t) , as above. In particular, we have ϕm := πum (um − u) ∈ Tum H 1,2 (IR/2π; S) . Note that, since n ∈ C 2 and since (um ) is bounded in H 1,2 (IR/2π; S), the sequence (ϕm ) is bounded in H 1,2 . Thus, ϕm , DE(um ) → 0. Moreover, since um u weakly in H 1,2 and uniformly, it follows that also ϕm 0 weakly in H 1,2 and uniformly. Consequently, we have 2π u˙ m ϕ˙ m dt o(1) = ϕm , DE(um ) =

0 2π

=

(u˙ m − u) ˙ ϕ˙ m dt + o(1)

0

d n(um ) n(um ) · (um − u) dt + o(1) dt 0 2π |u˙ m − u| ˙ 2 − (u˙ m − u) ˙ · n(um ) n(um ) · (u˙ m − u) ˙ dt + o(1) , = =

2π

|u˙ m − u| ˙ 2 − (u˙ m − u) ˙ ·

0

where o(1) → 0 as m → ∞. But n(u) · u˙ = 0 = n(um ) · u˙ m almost everywhere. Hence 2π |n(um )(u˙ m − u)| ˙ 2 dt = 0

2π

| (n(um ) − n(u)) u| ˙ 2 dt

0

≤ 2n(um ) − n(u)2∞ E(u) → 0 (m → ∞) , and um → u strongly. In order to construct a ﬂow-invariant family F we now proceed as follows: By assumption, there exists a C 3 -diﬀeomorphism Ψ : S → S 2 from S onto the standard sphere S 2 ⊂ IR3 . Let p: [− π2 , π2 ] → H 1,2 (IR/2π; S) be a 1-parameter family of closed curves u = p(θ) on S, such that p(± π2 ) ≡ p± are constant “curves” in S. Via Ψ we associate with p a map p˜ ∈ C 0 (S 2 ; S 2 ) by representing S 2 ∼ = [− π2 , π2 ] × IR/2π in polar coordinates (θ, φ), with {− π2 } × IR/2π, respectively { π2 } × IR/2π, collapsed to points. Then we let

92

Chapter II. Minimax Methods

p˜(θ, φ) = Ψ (p(θ)(φ)) . Consider now the collection π π π P = {p ∈ C 0 [− , ]; H 1,2 (IR/2π; S) ; p(± ) ≡ const. ∈ S} 2 2 2 and let F = {p ∈ P ; p˜ is homotopic to id S 2 } . Choosing for p(θ) the pre-image under Ψ of a family of equilateral circles covering S 2 , we ﬁnd that F = ∅. Also note that the map P p → p˜ ∈ C 0 (S 2 ; S 2 ) is continuous. Hence F is Φ-invariant under any homeomorphism Φ of H 1,2 (IR/2π; S) homotopic to the identity, and which maps constant maps to constant maps. Note that, in particular, any Φ which does not increase E will have this latter property.

S

p( θ )

Fig. 4.1. An admissible comparison path p ∈ P

Finally, by Theorem 4.2 β = inf sup E(u) p∈F u∈p

is critical. This almost completes the proof of Theorem 4.4. However, it remains to rule out the possibility that β = 0: the energy of trivial (constant) “closed geodesics”.

4. The Minimax Principle

93

4.5 Lemma. β > 0. Proof. There exists δ > 0 such that for any x at distance dist(x, S) ≤ δ from S there is a unique nearest neighbor π(x) ∈ S, characterized by |π(x) − x| = inf |x − y| , y∈S

and π(x) depends continuously on x. Moreover, π is C 2 if S is of class C 3 . By (4.1) there exists γ > 0 such that for u ∈ H 1,2 (IR/2π; S) with E(u) ≤ γ there holds (4.5)

diam(u) =

sup

0≤φ,φ ≤2π

|u(φ) − u(φ )| < δ .

Now suppose β < γ, and let p ∈ F be such that E(u) ≤ γ for any curve u = p(θ) ∈ p. By (4.5), if we ﬁx φ0 ∈ [0, 2π], we can continuously contract any such curve u to the constant curve u(φ0 ) in the δ-neighborhood of S by letting us (φ) = (1 − s)u(φ) + s u(φ0 ), 0 ≤ s ≤ 1 . Composing with π, we obtain a homotopy ps (θ, φ) = π (1 − s)p(θ)(φ) + sp(θ)(φ0 ) , 0 ≤ s ≤ 1, between p = p0 and a path p1 ∈ P consisting entirely of constant loops p1 (θ) ≡ p(θ)(φ0 ) for all θ. Composing ps with Ψ : S → S 2 , we also obtain a homotopy of p˜ = p˜0 ∼ id S 2 to the map p˜1 , given by p˜1 (θ, φ) = Ψ (p(θ)(φ0 )). But, letting p˜1,r (θ, φ) = p˜1 (rθ, φ) = Ψ p(rθ)(φ0 ) , 0 ≤ r ≤ 1 , we see that p˜1 – and hence also p˜ – is homotopic to a constant map, contrary to our choice of p ∈ F . 4.6 Notes. Birkhoﬀ’s result of 1917 and a later extension to spheres of arbitrary dimension (Birkhoﬀ [2]) mark the beginning of the calculus of variations in the large. A major advance then came with the celebrated work of LusternikSchnirelmann [1] in 1929 who – by variational techniques – established the existence of three geometrically distinct closed geodesics free of self-intersections on any compact surface of genus 0. (Detailed proofs were published by Lusternik [1] in 1947.) For recent developments in the theory of closed geodesics, see, for instance, Klingenberg [1], Bangert [1], and Franks [1].

94

Chapter II. Minimax Methods

5. Index Theory In most cases the topology of the space M where a functional E is deﬁned will be rather poor. However, if E is invariant under a compact group G acting on M , this may change drastically if we can pass to the quotient M/G with respect to the symmetry group. Often this space will have a richer topological structure which we may hope to exploit in order to obtain multiple critical points. However, in general the group G will not act freely on M and the quotient space will be singular, in particular, it will no longer be a manifold. Therefore the results outlined in the preceding sections cannot be applied. A nice way around this diﬃculty is to consider ﬂow-invariant families F in Theorem 4.2 which are also invariant under the group action. Since by Remark 3.5.(4◦ ) we may choose our pseudo-gradient ﬂows Φ to be equivariant if E is, this approach is promising. Moreover, at least for special kinds of group actions G, the topological complexity of the elements of such equivariant families can be easily measured or estimated in terms of an “index” which then may be used to distinguish diﬀerent critical points.

Krasnoselskii Genus The concept of an index theory is most easily explained for an even functional E on some Banach space V , with symmetry group G = ZZ2 = {id, −id}. Deﬁne A = {A ⊂ V ; A closed, A = −A} to be the class of closed symmetric subsets of V . 5.1 Definition. For A ∈ A, A = ∅, following Coﬀman [1], let 5 γ(A) =

inf m ; ∃h ∈ C 0 (A; IRm \ { 0}), h(−u) = −h(u) ∞,

if {..} = ∅, in particular, if A 0 ,

and deﬁne γ(∅) = 0. Note that for any A ∈ A, by the Tietze extension theorem, ˜ ∈ C 0 (V ; IRm ). any odd map h ∈ C 0 (A; IRm )may be extended to a map h 1 ˜ ˜ Letting h(u) = h(u) − h(−u) , the extension can be chosen to be symmetric. 2

γ(A) is called the Krasnoselskii genus of A. (The equivalence of Coﬀman’s deﬁnition above with Krasnoselskii’s [1] original deﬁnition – see also Krasnoselskii-Zabreiko [1; p. 385 ﬀ.] – was established by Rabinowitz [1; Lemma 3.6].) A notion of coindex with related properties was introduced by Connor-Floyd [1]. The notion of genus generalizes the notion of dimension of a linear space:

5. Index Theory

95

5.2 Proposition. For any bounded symmetric neighborhood Ω of the origin in IRm there holds: γ(∂Ω) = m. Proof. Trivially, γ(∂Ω) ≤ m. (Choose h = id.) Let γ(∂Ω) = k and let h ∈ C 0 (IRm ; IRk ) be an odd map such that h(∂Ω) 0. We may consider IRk ⊂ IRm . But then the topological degree of h: IRm → IRk ⊂ IRm on Ω with respect to 0 is well-deﬁned (see Deimling [1; Deﬁnition 1.2.3]). In fact, since h is odd, by the Borsuk-Ulam theorem (see Deimling [1; Theorem 1.4.1]) we have deg(h, Ω, 0) = 1 . Hence by continuity of the degree also deg(h, Ω, y) = 1 = 0 for y ∈ IR close to 0 and thus, by the solution property of the degree, h covers a neighborhood of the origin in IRm ; see Deimling [1; Theorem 1.3.1]. But then k = m, as claimed. m

Proposition 5.2 has a converse: 5.3 Proposition. Suppose A ⊂ V is a compact symmetric subset of a Hilbert space V with inner product (·, ·)V , and suppose γ(A) = m < ∞. Then A contains at least m mutually orthogonal vectors uk , 1 ≤ k ≤ m, (uk , ul )V = 0 (k = l). Proof. Let u1 , . . . , ul be a maximal set of mutually orthogonal vectors in A, and denote W = span {u1 , . . . , ul } ∼ = IRl , π: V → W orthogonal projection onto W . Then π(A) 0, and π deﬁnes an odd continuous map h = π|A : A → IRl \ {0}. By deﬁnition of γ(A) = m we conclude that l ≥ m, as claimed. Moreover, the genus has the following properties: 5.4 Proposition. Let A, A1 , A2 ∈ A, h ∈ C 0 (V ; V ) an odd map. Then the following hold: (1◦) γ(A) ≥ 0, γ(A) = 0 ⇔ A = ∅. (2◦) A1 ⊂ A2 ⇒ γ(A1 ) ≤ γ(A2 ). (3◦) γ(A1 ∪ A2) ≤ γ(A 1 ) + γ(A2 ).

(4◦) γ(A) ≤ γ h(A) . (5◦) If A ∈ A is compact and 0 ∈ A, then γ(A) < ∞ and there is a neighborhood N of A in V such that N ∈ A and γ(A) = γ(N ). That is, γ is a deﬁnite, monotone, sub-additive, supervariant and “continuous” map γ: A → IN0 ∪ {∞}. Proof. (1◦ ) follows by deﬁnition.

96

Chapter II. Minimax Methods

(2◦) If γ(A2 ) = ∞ we are done. Otherwise, suppose γ(A2 ) = m. By deﬁnition there exists h ∈ C 0 (A2 ; IRm \ {0}), h(−u) = −h(u). Restricting h to A1 yields an odd map h|A1 ∈ C 0 (A1 ; IRm \ {0}), whence γ(A1 ) ≤ γ(A2 ). (3◦) Again we may suppose that both γ(A1 ) = m1 , γ(A2 ) = m2 are ﬁnite, and we may let h1 , h2 be odd maps hi ∈ C 0 (Ai ; IRmi \ {0}), i = 1, 2, as in the deﬁnition of the genus. As noted above, we may extend h1 , h2 to odd maps hi ∈ C 0 (V ; IRmi ), i = 1, 2. But then letting h(u) = (h1 (u), h2 (u)) deﬁnes an odd map h ∈ C 0 (V ; IRm1 +m2 ) which does not vanish for u ∈ A1 ∪ A2 , and claim (3◦ ) follows. ˜ ∈ C 0 h(A); IRm \ {0} induces an odd map h ˜◦h ∈ (4◦) Any odd map h C 0 (A; IRm \ {0}), and (4◦ ) is immediate. (5◦) If A is compact and 0 ∈ A there is ρ > 0 such that A ∩ Bρ (0) = ∅. ˜ of A admits a ﬁnite sub-cover The cover Bρ (u) = Bρ (u) ∪ Bρ (−u) u∈A

˜ρ (u1 ), . . . , B ˜ ρ (um )}. Let {ϕj }1≤j≤m be a partition of unity on A sub{B ˜ ρ (uj )}1≤j≤m ; that is, let ϕj ∈ C 0 B ˜ρ (uj ) with support in ordinate to {B m ˜ρ (uj ) satisfy 0 ≤ ϕj ≤ 1, B ϕ (u) = 1 for all u ∈ A. Replacing ϕj j=1 j by ϕj (u) = 12 ϕj (u) + ϕj (−u) , if necessary, we may assume that ϕj is even, 1 ≤ j ≤ m. By choice of ρ, for any j the neighborhoods Bρ (uj ), Bρ (−uj ) are disjoint. Hence the map h: V → IRm with jth component ⎧ if u ∈ Bρ (uj ) , ⎨ ϕj (u), hj (u) = −ϕj (u), if u ∈ Bρ (−uj ) , ⎩ 0, else , is continuous, odd, and does not vanish on A. This shows that γ(A) < ∞. Finally, assume that A is compact, 0 ∈ A, γ(A) = m < ∞ and let h ∈ C 0 (A; IRm \ {0}) be as in the deﬁnition of γ(A). We may assume h ∈ C 0 (V ; IRm ). Moreover, A being compact, also h(A) is compact, and there ˜ of h(A) whose closure is compactly exists a symmetric open neighborhood N ˜ ), by construction h(N ) 0 and contained in IRm \ {0}. Choosing N = h−1 (N γ(N ) ≤ m. On the other hand, A ⊂ N . Hence γ(N ) = γ(A) by monotonicity of γ, property (2◦ ). 5.5 Observation. It is easy to see that if A is a ﬁnite collection of antipodal pairs ui , −ui (ui = 0), then γ(A) = 1. Minimax Principles for Even Functionals Suppose E is a functional of class C 1 on a closed symmetric C 1,1 -submanifold M of a Banach space V and satisﬁes (P.-S.). Moreover, suppose that E is even, that is, E(u) = E(−u) for all u. Also let A be as above. Then for any k ≤ γ(M ) ≤ ∞ by Proposition 5.4(4◦ ) the family

5. Index Theory

97

Fk = {A ∈ A ; A ⊂ M, γ(A) ≥ k} is invariant under any odd and continuous map and non-empty. Hence, analogous to Theorem 4.2, for any k ≤ γ(M ), if βk = inf

sup E(u)

A∈Fk u∈A

is ﬁnite, then βk is a critical value of E; see Theorem 5.7 below. To see what happens when βk and βk+1 coincide for some k, it is instructive to compare this result with the well-known Courant-Fischer minimax principle for linear eigenvalue problems. Recall that on IRn with scalar product (·, ·) the k-th eigenvalue of a symmetric linear map K: IRn → IRn is given by λk =

min

V ⊂V, dim V =k

max (Ku, u) .

u∈V u=1

Translated into the above setting, we may likewise determine λk by considering the functional E(u) = (Ku, u) on the unit sphere M = S n−1 and computing βk as above. Trivially, E satisﬁes (P.-S.); moreover, it is easy to see that βk = λk for all k. (By Proposition 5.2 the inequality βk ≤ λk is immediate. The reverse inequality follows by Proposition 5.3 and linearity of K.) In the linear case, now, it is clear that if successive eigenvalues λk = λk+1 = . . . = λk+l−1 = λ coincide, then K has an l-dimensional eigenspace of eigenvectors u ∈ V satisfying Ku = λu. Is there a similar result in the nonlinear setting? Actually, there is. For this we again assume that E ∈ C 1 (M ) is an even functional on a closed, symmetric C 1,1 -submanifold M ⊂ V \ {0}, satisfying (P.-S.). Let βk , k ≤ γ(M ), be deﬁned as above. 5.6 Lemma. Suppose for some k, l there holds −∞ < βk = βk+1 = . . . = βk+l−1 = β < ∞ . Then γ(Kβ ) ≥ l. By Observation 5.5, in particular, if l > 1, Kβ is inﬁnite. Proof. By (P.-S.) the set Kβ is compact and symmetric. Hence γ(Kβ ) is welldeﬁned and by Proposition 5.4(5◦ ) there exists a symmetric neighborhood N of Kβ in M such that γ(N) = γ(Kβ ). For ε = 1, N , and β as above let ε > 0 and Φ be determined according to Theorem 3.11. We may assume Φ is odd. Choose A ⊂ M such that γ(A) ≥ k + l − 1 and E(u) < β + ε for u ∈ A. Let Φ(A, 1) = A˜ ∈ A. By property (3◦ ) of Φ in Theorem 3.11 A˜ ⊂ (Eβ−ε ∪ N ) . Moreover, by deﬁnition of β = βk it follows that γ Eβ−ε < k .

98

Chapter II. Minimax Methods

Thus, by Proposition 5.4(2◦ )–5.4(4◦ )we have γ(N ) ≥ γ Eβ−ε ∪ N − γ Eβ−ε ˜ − k ≥ γ(A) − k > γ(A) ≥k+l−1−k =l−1 ; that is, γ(N ) = γ(Kβ ) ≥ l, as claimed. In consequence, we note 5.7 Theorem. Suppose E ∈ C 1 (M ) is an even functional on a complete symmetric C 1,1 -manifold M ⊂ V \ {0} in some Banach space V . Also suppose E satisﬁes (P.-S.) and is bounded from below on M . Let γˆ (M ) = sup{γ(K) ; K ⊂ M compact and symmetric}. Then the functional E possesses at least γˆ (M ) ≤ ∞ pairs of critical points. Remarks. Note that the deﬁnition of γˆ(M ) assures that for k ≤ γˆ (M ) the numbers βk are ﬁnite. Completeness of M can be replaced by the assumption that the ﬂow deﬁned by any pseudo-gradient vector ﬁeld on M exists for all positive time. Applications to Semilinear Elliptic Problems As a particular case, Theorem 5.7 includes the following classical result of Lusternik-Schnirelmann [2]: Any even function E ∈ C 1 (IRn ) admits at least n distinct pairs of critical points when restricted to S n−1 . In inﬁnite dimensions, Theorem 5.7 and suitable variants of it have been applied to the solution of nonlinear partial diﬀerential equations and nonlinear eigenvalue problems with a ZZ2 -symmetry. See for instance Amann [1], Clark [1], Coﬀman [1], Hempel [1], Rabinowitz [7], Thews [1], and the surveys and lecture notes by Browder [2], Rabinowitz [7], [11]. Here we present only a simple example of this kind for which we return to the setting of problem (I.2.1) (I.2.3)

−Δu + λu = |u|p−2 u u=0

in Ω , on ∂Ω ,

considered earlier. This time, however, we also admit solutions of varying sign. 5.8 Theorem. Let Ω be a bounded domain in IRn , and let p > 2; if n ≥ 3 2n . Then for any λ ≥ 0 problem (I.2.1), (I.2.3) we assume in addition p < n−2 admits inﬁnitely many distinct pairs of solutions. Proof. By Theorem 5.7 the even functional

5. Index Theory

E(u) =

1 2

99

|∇u|2 + λ|u|2 dx

Ω

admits inﬁnitely many distinct pairs of critical points on the sphere S = {u ∈ H01,2 ; uLp = 1}, for any λ ≥ 0. Scaling suitably, we obtain inﬁnitely many distinct pairs of solutions for (I.2.1), (I.2.3). General Index Theories The concept of index can be generalized. Our presentation is based on Rabinowitz [11]. Related material can also be found in the recent monograph by Bartsch [1]. Suppose M is a complete C 1,1 -Finsler manifold with a compact group action G. Let A = {A ⊂ M ; A is closed, g(A) = A for all g ∈ G} be the set of G-invariant subsets of M , and let Γ = {h ∈ C 0 (M ; M ) ; h ◦ g = g ◦ h for all g ∈ G} be the class of G-equivariant mappings of M . (Since our main objective is that the ﬂow Φ( · , t) constructed in Theorem 3.11 be in Γ , we might also restrict Γ to the class of G-equivariant homeomorphisms of M .) Finally, if G = {id}, denote Fix G = u ∈ M ; gu = u for all g ∈ G the set of ﬁxed points of G. 5.9 Definition. An index for (G, A, Γ ) is a mapping i: A → IN0 ∪ {∞} such that for all A, B ∈ A, h ∈ Γ there holds (1◦) (deﬁniteness:) i(A) ≥ 0, i(A) = 0 ⇐⇒ A = ∅. (2◦) (monotonicity:) A ⊂ B ⇒ i(A) ≤ i(B). (3◦) (sub-additivity:) i(A ∪ B)≤ i(A) + i(B). (4◦) (supervariance:) i(A) ≤ i h(A) .

(5◦) (continuity:) If A is compact and A ∩ F ix G = ∅, then i(A) < ∞ and there is a G-invariant neighborhood N of A such that i(N ) = i(A). ◦ (6 ) (normalization): If u ∈ F ix G, then i g∈G gu = 1.

5.10 Remarks and examples. (1◦ ) If A ∈ A and A ∩ F ix G = ∅ then i(A) = supB∈A i(B); indeed, by monotonicity, for u0 ∈ A∩F ix G there holds i ({uo }) ≤ i(A) ≤ supB∈A i(B). On the other hand, for any B ∈ A the map h: B → {u0 }, given by h(u) = u0 for all u ∈ B, is continuous and equivariant, whence i(B) ≤ i ({u0 }) by supervariance of the index. Hence, in general nothing will be lost if we deﬁne i(A) = ∞ for A ∈ A such that A ∩ F ix G = ∅.

100

Chapter II. Minimax Methods

(2◦) By Example 5.1 the Krasnoselskii genus γ is an index for G = {id, −id}, the class A of closed, symmetric subsets, and Γ the family of odd, continuous maps. Analogous to Theorem 5.7 we have the following general existence result for variational problems that admit an index theory. 5.11 Theorem. Let E ∈ C 1 (M ) be a functional on a complete C 1,1 -Finsler manifold M and suppose E is bounded from below and satisﬁes (P.-S.). Suppose G is a compact group acting on M without ﬁxed points and let A be the set of closed G-invariant subsets of M , Γ be the group of G-equivariant homeomorphisms of M . Suppose i is an index for (G, A, Γ ), and let ˆi(M ) = sup{i(K) ; K ⊂ M is compact and G − invariant} ≤ ∞ . Then E admits at least ˆi(M ) critical points which are distinct modulo G. The proof is the same as that of Theorem 5.7 and Lemma 5.6. Again note that completeness of M can be replaced by the assumption that any pseudo-gradient ﬂow on M is complete in forward time. Lusternik-Schnirelman Category The concept of category was introduced by Lusternik-Schnirelmann [2]. This notion, in fact, is the ﬁrst example of an index theory (in the above sense) in the mathematical literature. 5.12 Definition. Let M be a topological space and consider a closed subset A ⊂ M . We say that A has category k relative to M (catM (A) = k), if A is covered by k closed sets Aj , 1 ≤ j ≤ k, which are contractible in M , and if k is minimal with this property. If no such ﬁnite covering exists, we let catM (A) = ∞. Moreover, we deﬁne catM (∅) = 0. This notion ﬁts in the frame of Deﬁnition 5.9 if we let G = {id} A = {A ⊂ M ; A closed }, Γ = {h ∈ C 0 (M ; M ) ; h is a homeomorphism }. Then we have 5.13 Proposition. catM is an index for (G, A, Γ ). Proof. (1◦ )–(3◦ ) of Deﬁnition 5.9 are immediate. (4◦ ) is also clear, since a homeomorphism h preserves the topological properties of any sets Aj covering A. (5◦ ) Any open cover of a compact set A by open sets O whose closure is contractible has a ﬁnite subcover {Oj , 1 ≤ j ≤ k}. Set N = Oj . (6◦ ) is obvious.

5. Index Theory

101

5.14 Categories of some standard sets. (1◦ ) If M = T m = IRm /ZZm is the m-torus, then catT m (T m ) = m + 1, see Lusternik-Schnirelman [2] or Schwartz [2; Lemma 5.15, p. 161]. Thus, any functional E ∈ C 1 (T m ) possesses at least m + 1 distinct critical points. In particular, if m = 2, any C 1 -functional on the standard torus, besides an absolute minimum and maximum, must possess at least one additional critical point. (2◦) For the m-sphere S m ⊂ IRm+1 we have catS m (S m ) = 2. (Take A1 , A2 slightly overlapping northern and southern hemispheres.) (3◦) For the unit sphere S in an inﬁnite dimensional Banach space we have catS (S) = 1. (S is contractible in itself.) (4◦) For real or complex m-dimensional projective space P m we have catP m (P m ) = m + 1 (m ≤ ∞). Since real projective P m = S m /ZZ2 , we may ask whether, in the presence of a ZZ2 -symmetry u → −u, the category and Krasnoselskii genus of symmetric sets are always related as in the above example 5.14(4◦ ). This is indeed the case, see (Rabinowitz [1; Theorem 3.7]): 5.15 Proposition. Suppose A ⊂ IRm \ {0} is compact and symmetric, and let ˜ A˜ = A/ZZ2 with antipodal points collapsed. Then γ(A) = catIRm \{0}/ZZ2 (A). Using the notion of category, results in the spirit of Theorem 5.8 have been established by Browder [1], [3] and Schwartz [1], for example. With index theories oﬀering a very convenient means to characterize diﬀerent critical points of functionals possessing certain symmetries, it is not surprising that, besides the classical examples treated above, a variety of other index theories have been developed. See the papers by Fadell-Husseini [1], FadellHusseini-Rabinowitz [1], Fadell-Rabinowitz [1] on cohomological index theories – a very early paper in this regard is Yang [1]. Relative or pseudo-indices were introduced by Benci [3] and Bartolo-Benci-Fortunato [1]. For our ﬁnal model problem in this section it will suﬃce to consider the S 1 -index of Benci [2] as another particular case. A Geometrical S 1 -Index If M is a complete C 1,1 -Finsler manifold with an S 1 -action (in particular, if M is a complex Hilbert space with S 1 = {eiφ ; 0 ≤ φ ≤ 2π} acting through scalar multiplication) we may deﬁne an index for this action as follows; see Benci [2]. 5.16 Definition. Let A be the family of closed, S 1 -invariant subsets of M , and Γ the family of S 1 -equivariant maps (or homeomorphisms). For A = ∅, deﬁne ⎧ ⎨ inf m ; ∃h ∈ C 0 A; Cm \ {0} , l ∈ IN : τ (A) = if {. . .} = ∅, h ◦ g = g l ◦ h for all g ∈ S 1 , ⎩ ∞, if {. . .} = ∅, and let τ (∅) = 0. (Note the similarity with the Krasnoselskii index γ.) Here, S 1 acts on Cm by component-wise complex multiplication.

102

Chapter II. Minimax Methods

5.17 Proposition. τ is an index for (S 1 , A, Γ ). Proof. It is easy to see that τ satisﬁes properties (1◦ ), (2◦ ), and (4◦ ) of Definition 5.9. To see (3◦ ) we may assume that τ (Ai ) = mi < ∞, i = 1, 2, and we may choose hi , li as in the deﬁnition of τ such that hi ∈ C 0 (Ai ; Cmi \ {0}) satisﬁes hi ◦ g = g li ◦ hi for all g ∈ S 1 , i = 1, 2. Extending hi to M and averaging ˜ i (u) = g −li hi (gu) dg h S1

with respect to an invariant measure (arc-length) on S 1 , we may assume that hi ∈ C 0 (M ; Cm ), i = 1, 2. But then the map l2 l1 , h(u) = h1 (u) , h2 (u) l where for (z1 , . . . , zm ) ∈ Cm , l ∈ IN we let (z1 , . . . , zm )l := (z1l , . . . , zm ), deﬁnes a map h ∈ C 0 A1 ∪ A2 ; Cm1 +m2 \ {0} l l2 l l1 1 2 such that h ◦ g = g ◦ h1 , g ◦ h2 = g l ◦ h for all g ∈ S 1 with l = l1 l2 .

To see (6◦ ), for an element u0 ∈ F ix (S 1 ) let G0 = {g ∈ S 1 ; gu0 = u0 } be the subgroup of S 1 ﬁxing u0 . Since u0 ∈ F ix (S 1 ), G0 is discrete, hence represented by G0 = e2πik/l ; 0 ≤ k < l for some l ∈ IN. For u = gu0 now let h(u) = g l ∈ S 1 ⊂ C \ {0}. Then h is well-deﬁned and continuous along the S 1 -orbit S 1 u0 = {gu0 ; g ∈ S 1 } of u0 . Extending h equivariantly, we see that τ (S 1 u0 ) = 1, which proves (6◦ ). Finally, to see (5◦ ), suppose A is S 1 -invariant, compact, and A∩Fix(S 1 ) = ◦ ∅. For any u0 ∈ A let h be constructed as in the proof of (6 ) and let O(u0 ) be an S 1 -invariant neighborhood of S 1 u0 such that h O(u0 ) 0. By compactness of A, ﬁnitely many such neighborhoods {O(ui )}1≤i≤m cover A, whence τ (A) < ∞ by sub-additivity. The remainder of the proof of (5◦ ) is the same as in Proposition 5.4. As in the case of the Krasnoselskii genus, Proposition 5.2, the S 1 -index may be interpreted as a generalized dimension of a closed S 1 -invariant set; see Benci [2; Proposition 2.6]. However, we will not pursue this. Instead, we observe that in the case of a free S 1 -action on a manifold M , a simpler variant of Benci’s S 1 -index can be deﬁned as follows. (Recall that a group G acts freely on a manifold M if only the identity element in G ﬁxes points in M .)

5. Index Theory

103

5.16 Definition. Suppose S 1 acts freely on a manifold M . Let A be the family of closed, S 1 -invariant subsets of M , and Γ the family of S 1 -equivariant maps (or homeomorphisms). For A = ∅, deﬁne ⎧ ⎨ inf m ; ∃h ∈ C 0 A; Cm \ {0} : τ˜(A) = h ◦ g = g ◦ h for all g ∈ S 1 , ⎩ ∞, if {. . .} = ∅, and let τ˜(∅) = 0. The proof of Proposition 5.17 may be carried over easily to see that τ˜ is an index for (S 1 , A, Γ ). Multiple Periodic Orbits of Hamiltonian Systems As an application, we present the following theorem on the existence of “many” periodic solutions of Hamiltonian systems, due to Ekeland and Lasry [1]: 5.18 Theorem. Suppose H ∈ C 1 (IR2n ; IR), and for some β > 0 assume that C = {x ∈ IR2n ; H(x) ≤ β} is strictly convex, with boundary S = {x ∈ IR2n ; H(x) = β} satisfying x · ∇H(x) > 0 for x ∈ S. Suppose that for numbers r, R > 0 with √ r
x˙ = J ∇H(x)

on S. (J is deﬁned on p. 60.) Theorem 5.18 provides a “global” analogue of a result by Weinstein [1] on the existence of periodic orbits of Hamiltonian systems near an equilibrium. Further extensions and generalizations of Theorem 5.18 were given by AmbrosettiMancini [2] and Berestycki-Lasry-Mancini-Ruf [1] who allow for energy surfaces “pinched” between ellipsoids rather than spheres. Moreover, Ekeland-Lassoued [1] have been able to show that any strictly convex energy surface carries at least two distinct periodic orbits of (5.1). It is conjectured that a result like Theorem 5.18 holds true in general on such surfaces; the proof of this conjecture, however, remains open. Proof of Theorem 5.18. We follow Ambrosetti and Mancini [2]. As observed in Section I.6 we may assume that (5.2)

H(sx) = sq H(x)

104

Chapter II. Minimax Methods

is homogeneous of degree q, 1 < q < 2, and strictly convex. Moreover, dividing H by βq we may assume that β = 1q . By our assumption on S and (5.2), ﬁnally, we have (5.3)

1 1 |x|q ≤ H(x) ≤ q |x|q . qRq qr

Let H ∗ be the Legendre-Fenchel transform of H as in I.6. Recall that H ∗ ∈ C 1 , q > 2; moreover, (5.3) translates into and H ∗ (sy) = sp H ∗ (y) with p = q−1 (5.4)

1 1 p p r |y| ≤ H ∗ (y) ≤ Rp |y|p . p p

Also let K be the integral operator

t

J y dt

(Ky)(t) = 0

on

Lp0 = y ∈ Lploc IR; IR2n ; y(t + 2π) = y(t),

2π

y dt = 0 .

0

Then – as described in detail in Section I.6, following Equations (I.6.5 ), (I.6.6 ) – a function y ∈ Lp0 \ {0} solves the Equation (I.6.7), that is, the equation (5.5)

∇H ∗ (y) − Ky = x0

for some x0 ∈ IR2n , if and only if the function x = Ky + x0 ∈ C 1 ([0, 2π]; IR2n ) solves (5.1) with H x(t) =: h/q > 0, which in turn implies that x ˜(t) = ˜(t) ≡ β = 1q . h−1/q x h2/q−1 t solves (5.1) with H x Suppose we can exhibit n distinct solutions yk of (5.5) with minimal period 2π corresponding to distinct solutions xk of (5.1) with energies H xk (t) = hk . ˜j = x ˜k (since the solutions xk are Then either hj = hk and the corresponding x ˜j , x ˜k will have diﬀerent minimal period and hence distinct). Or hj = hk and x be distinct. In any event we will have achieved the proof of the theorem. This is the strategy that we now follow. Denote E: Lp0 → IR the dual variational integral corresponding to (5.5), given by 2π 1 E(y) = H ∗ (y) − y, Ky dt . 2 0 Note that we have an S 1 -action on Lp0 , via (τ, y) → yτ (t) = y(t + τ ) ,

for all γ = eiτ ∈ S 1 .

This action leaves E invariant. Moreover, y has minimal period 2π if and only if yτ = y precisely for τ ∈ 2πZZ. Denote m = inf E(y) ; y ∈ Lp0 and let

5. Index Theory

105

m∗ = inf E(y) ; y ∈ Lp0 , ∃k ∈ IN, k ≥ 2 : yτ = y for τ = 2π/k . Observe that, since p > 2 and since the spectrum of K contains positive eigenvalues, we have m ≤ m∗ < 0. Also note that the S 1 -action will be free on the set M = {y ∈ Lp0 ; m ≤ E(y) < m∗ } . In particular, any y ∈ M will have minimal period 2π. Finally, E satisﬁes condition (P.-S.). Indeed, since E is coercive on Lp0 , we may assume that any (P.-S.)-sequence (ym ) for E is bounded in Lp0 . Thus ym y weakly in Lp0 and Kym → Ky strongly in Lq [0, 2π]; IR2n . Recall from Section I.6 that by strict convexity and homogeneity of H ∗ the diﬀerential ∇H ∗ is strongly monotone. That is, we have o(1)ym − yLp = ym − y, DE(ym) − DE(y) 2π ym − y, ∇H ∗ (ym ) − ∇H ∗ (y) − ym − y, Kym − Ky dt = 0 ≥ α ym − yLp ym − yLp − o(1) , where o(1) → 0 as m → ∞, with a non-negative, non-decreasing function α that vanishes only at 0. If follows that ym → y strongly in Lp0 , as claimed. Thus, by Theorem 5.11, the proof will be complete if we can show that the (simpliﬁed) S 1 -index of a suitable compact S 1 -invariant subset of M is ≥ n. (From the deﬁnition of M it is clear that any pseudo-gradient ﬂow on M will be complete in forward time.) Note that since E is weakly lower semi-continuous and coercive on Lp0 and since in both cases the set of comparison functions is weakly closed, m and m∗ will be attained in their corresponding classes. Let y ∗ ∈ Lp0 satisfy E(y ∗ ) = m∗ , yτ∗ = y ∗ for some τ ∈ 2πZZ . By minimality of y ∗ ∗

∗

2π

y , DE(y ) = p

∗

∗

H (y ) dt −

0

2π

y ∗ , Ky ∗ dt = 0 ,

0

whence in particular m∗ = We may assume that τ = comparison function

2π k

1 1 − p 2

y ∗ , Ky ∗ dt .

0

for some k ∈ IN, k > 1. Hence we obtain as

t y(t) = y ∈ Lp0 , k ∗

and

2π

106

Chapter II. Minimax Methods

m ≤ inf E(sy) = inf sp s>0

But

s>0

2π

0

H ∗ (y) dt =

2π

0

2π

s2 H (y) dt − 2 ∗

1 p

H ∗ (y ∗ ) dt =

0

2π

2π

y, Ky dt

.

0

y ∗ , Ky ∗ dt ,

0

while

2π

(5.6)

y, Ky dt = k

0

2π

y ∗ , Ky ∗ dt .

0

Hence

2π sp s2 − k y ∗ , Ky ∗ dt p 2 0 2π 1 1 − y ∗ , Ky ∗ dt p 2 0

m ≤ inf

s>0

(5.7)

p

= k p−2 p

≤ 2 p−2 m∗ < 0 . To obtain a lower bound on m, let y ∈ Lp0 satisfy E(y) = m . Then

2π

y, DE(y) = p

(5.8)

H ∗ (y) dt −

0

and hence (5.9)

m=

1 1 − p 2

0

2π

2π

y, Ky dt = 0 ,

0

p 2π ∗ y, Ky dt = 1 − H (y) dt . 2 0

Note that by (5.4) for z ∈ Lp0 we have p/2 2π 2π rp rp 1 p 2 (5.10) H (z) dt ≥ |z| dx ≥ |z| dx . 2πp 0 p 2π 0 0

2π 2 1 Let Σ = {z ∈ Lp0 ; 2π |z| dt = 1} and denote 0

2π z, Kz dt ; z ∈ Σ > 0 . b = sup 1 2π

2π

∗

0

Then, if we let y = λz, z ∈ Σ, by (5.8), (5.10) we have 2π 2π z, Kz dt = p H ∗ (z) dt λ2−p b ≥ λ2−p (5.11) 0 0 ≥ 2πr p .

5. Index Theory

107

Since p > 2 this implies a bound from above for λ. By estimating 2π 1 1 1 1 2 − − m= z, Kz dt ≥ λ λ2 b p 2 p 2 0 (5.12) 1 1 − ≥ (2π)−2/(p−2) bp/(p−2) r −2p/(p−2) =: c0 r −2p/(p−2) p 2 with a constant c0 < 0, the latter translates into a lower bound for m, and hence for m∗ , by (5.7). Now let

Σn = z(t) = eJ t (ξ, η) = (ξ cos t − η sin t, ξ sin t + η cos t) ; 2n 2 2 (ξ, η) ∈ IR with |ξ| + |η| = 1 ⊂ Σ . Clearly Σn is S 1 -invariant. Moreover, note that for z ∈ Σn we have H ∗ z(t) ≤ Rp /p for all t, whence 2π 1 (5.13) H ∗ (z) dt ≤ Rp /p . 2π 0 Now, we have Lemma 5.19.

2π 0

z, Kz dt = b for any z ∈ Σn .

Postponing the proof of Lemma 5.19, we conclude the proof of Theorem 5.18 as follows. For z ∈ Σn let λ = λ(z) > 0 satisfy 2π 2π p ∗ 2 H (z) dt − λ z, Kz dt = 0 . (5.14) pλ 0

0

By (5.10) the map z → λ(z)z is an S -equivariant C 1 -embedding of Σn into ˜ n diﬀeomorphic to Σn by radial Lp0 , mapping Σn onto an S 1 -invariant set Σ 1 ˜ n ) ≥ τ˜(Σn ). projection; in particular, the (simpliﬁed) S -index τ˜(Σ From (5.13), (5.14) and Lemma 5.19, as in (5.11) we now obtain that for z ∈ Σn , λ = λ(z) there holds 2π 2π z, Kz dt = p H ∗ (z) dt ≤ 2πRp . λ2−p b = λ2−p 1

0

0

Hence analogous to (5.12) we obtain sup E(y) ≤ c0 R−2p/p−2 < 0 .

˜n y∈Σ

Since by assumption R < sup E(y) < ˜n y∈Σ

√

2 r, and in view of (5.7), this implies that

1 1 c0 r −2p/(p−2) ≤ p/p−2 m ≤ m∗ ; 2p/(p−2) 2

108

Chapter II. Minimax Methods

˜ n ⊂ M . Hence the proof of Theorem 5.18 is complete if we show that that is, Σ τ˜(Σn ) ≥ n. But any S 1 -equivariant map h: Σn → Cm \{0}with h◦g = g ◦h induces an 2n−1 = (ξ, η) ∈ IR2n ; |ξ|2 +|η|2 = 1 into Cm \{0} ∼ odd map of S = IR2m \{0}, given by (ξ, η) → h eJ t (ξ, η) . By Proposition 5.2 we conclude that 2m ≥ 2n, whence τ˜(Σn ) ≥ n. (Since the map z = z(t) = eJ t (ξ, η) → z(0) = (ξ, η) ∈ IR2n \ {0} ∼ = Cn \ {0} is S 1 -equivariant, we actually have equality τ˜(Σn ) = n.) Proof of Lemma 5.19. Since K is compact there exists z ∈ Σ satisfying 2π z, Kz dt = b . 0

By (5.6), z must have minimal period 2π. Moreover, z satisﬁes Kz + x0 = λz

2π 1 for some λ ∈ IR, where x0 = − 2π 0 K(z) dt. Setting x = Kz + x0 , equivalently, x solves λx˙ = J x . Hence, |x| ≡ const. Moreover, the properties of z imply that x = 0 and that x has minimal period 2π. A scalar multiple of x (and hence also a scalar multiple of z) thus belong to Σn .

2π By S 1 -invariance of I(z) = 0 z, Kz dt and invariance of I under rotations in IR2n , if I(z) = b for some z ∈ Σn , it follows that I(z) = b for all z ∈ Σn . The proof is complete.

6. The Mountain Pass Lemma and its Variants The minimax principle and its variants essentially cover all possibilities how existence results for saddle points can be drawn from information about the topology of the sub-level sets of a functional E. However, unless the domain of E itself has a rich topology, ﬁnding the right notion of ﬂow-invariant family may be quite tiresome. Fortunately, there are existence results for saddle points tailor-made for applications. These are the famous (inﬁnite dimensional) mountain pass lemma and its variants, due to Ambrosetti and Rabinowitz [1]. The simplest form of these results reads as follows.

6. The Mountain Pass Lemma and its Variants

109

6.1 Theorem. Suppose E ∈ C 1 (V ) satisﬁes (P.-S.). Assume that (1◦) E(0) = 0 ; (2◦) ∃ρ > 0, α > 0 : u = ρ ⇒ E(u) ≥ α; (3◦) ∃u1 ∈ V : u1 ≥ ρ and E(u1 ) < α. Deﬁne P = p ∈ C 0 ([0, 1]; V ) ; p(0) = 0, p(1) = u1 . Then β = inf sup E(u) ≥ α p∈P u∈p

is a critical value.

E

M 0

u1 Fig. 6.1. On the mountain pass lemma of Ambrosetti and Rabinowitz

Proof. Suppose by contradiction that Kβ = ∅. For ε = min{α, α − E(u1 )} and N = ∅ determine ε > 0 and a deformation Φ as in Theorem 3.4. By deﬁnition of β, there exists p ∈ P such that sup E(u) < β + ε . u∈p

Consider p1 = Φ(p, 1). Note that by choice of ε the deformation Φ(·, 1) leaves u0 = 0 and u1 ﬁxed. Hence p1 ∈ P . Moreover, Φ(Eβ+ε , 1) ⊂ Eβ−ε ; therefore sup E(u) = sup E (Φ(u, 1)) < β − ε ,

u∈p1

u∈p

by choice of p. But this contradicts the deﬁnition of β, and the proof is complete. Remark. Observe that by Remark 3.5(3◦ ) the proof, and hence the assertion of Theorem 6.1, remain true if we only require (P.-S.) at the level β.

110

Chapter II. Minimax Methods

Applications to Semilinear Elliptic Boundary Value Problems Theorem 6.1 permits an alternative proof of Theorem I.2.1. However, Theorem 6.1 can also be applied to more general problems of the type (6.1) (6.2)

−Δu = g(·, u) u=0

in Ω , on ∂Ω ,

which cannot be solved by a constrained minimization method. 6.2 Theorem. Let Ω be a smooth, bounded domain in IRn , n ≥ 3, and let g: Ω × u IR → IR be a Carath´eodory function with primitive G(x, u) = 0 g(x, v) dv. Suppose that the following conditions hold: (1◦) g(x, 0) = 0 and lim supu→0 g(x,u) ≤ 0, uniformly in x ∈ Ω; u 2n , C: |g(x, u)| ≤ C 1+|u|p−1 , for almost every x ∈ Ω, u ∈ IR; (2◦) ∃p < 2∗ = n−2 (3◦) ∃q > 2, R0 : 0 < q G(x, u) ≤ g(x, u) u, for almost every x ∈ Ω, if |u| ≥ R0 . Then problem (6.1, (6.2) admits non-trivial solutions u+ ≥ 0 ≥ u− . If, in addition, g is H¨ older continuous in both variables, then u+ > 0 > u− in Ω.

u Remark. Here again, G(x, u) = 0 g(x, v) dv denotes a primitive of g. An analogous result is valid for n = 2. However, notation is simpler if we consider only n ≥ 3. Wang, Z.Q. [1], under (essentially) the assumptions of Theorem 6.2, recently has established the existence of a third non-trivial solution. (Note that by (1◦ ) problem (6.1), (6.2) always admits the solution u = 0.) Proof. Problem (6.1), (6.2) corresponds to the Euler-Lagrange equation of the functional 1 |∇u|2 dx − G(x, u) dx E(u) = 2 Ω Ω on the space H01,2 (Ω). Assumption (2◦ ) implies that E is of class C 1 . To see that E satisﬁes (P.-S.) , ﬁrst note that by (2◦ ) the map u → g(·, u) takes bounded sets in Lp (Ω) into bounded sets in Lp/(p−1) (Ω) ⊂ H −1 (Ω). 2n by Rellich’s theorem the space H01,2 (Ω) Therefore, and since for p < n−2 embeds into Lp (Ω) compactly, the map K: H01,2 (Ω) → H −1 (Ω), given by K(u) = g(·, u), is compact. Since DE is of the form DE(u) = −Δu − g(·, u), by Proposition 2.2 it suﬃces to show that any (P.-S.)-sequence (um ) for E is bounded in H01,2 (Ω). Let (um ) be a (P.-S.)-sequence. Then we obtain

(6.3)

C + o(1)um H 1,2 ≥ q E(um ) − um , DE(um ) 0 q−2 2 g(x, um )um − qG(x, um ) dx |∇um | dx + = 2 Ω Ω q−2 2 n ≥ um H 1,2 + L (Ω) · ess inf g(x, v)v − qG(x, v) , 0 x∈Ω,v∈IR 2

where o(1) → 0 (m → ∞).

6. The Mountain Pass Lemma and its Variants

111

But by (2◦ ) and (3◦ ) the last term is ﬁnite and the desired conclusion follows. Next observe that E(0) = 0; moreover, by (1◦ ), (2◦ ) for any ε > 0 there exists C(ε) such that G(x, u) ≤ ε|u|2 + C(ε)|u|p for all u ∈ IR and almost every x ∈ Ω. It follows that 1 E(u) ≥ |∇u|2 dx − ε |u|2 dx − C(ε) |u|p dx 2 Ω Ω Ω 1 ε p−2 2 ≥ − C(ε)uH 1,2 uH 1,2 ≥ α > 0 , − 0 2 λ1 0 provided uH 1,2 = ρ is suﬃciently small. Here λ1 denotes the ﬁrst eigenvalue 0 of −Δ in Ω with homogeneous Dirichlet boundary conditions, given by the Rayleigh-Ritz quotient (I.2.4). Moreover, we have used the Sobolev embedding H01,2 (Ω) → Lp (Ω); see Theorem A.5 of Appendix A. Finally, condition (3◦ ) can be restated as a diﬀerential inequality for the function G, of the form u|u|q

d −q |u| G(x, u) ≥ 0 , for |u| ≥ R0 . du

Upon integration we infer that for |u| ≥ R0 we have G(x, u) ≥ γ0 (x)|u|q with γ0 (x) = R0−q min G(x, R0 ), G(x, −R0 ) > 0. Hence, if u ∈ H01,2 (Ω) does not vanish identically, we obtain that with constants C(u), c(u) > 0 there holds λ2 |∇u|2 dx − G(x, λu) dx E(λu) = 2 Ω Ω (6.5) ≤ C(u) λ2 − c(u)λq + Ln (Ω) ess sup |G(x, v)| ,

(6.4)

x∈Ω,|v|≤R0

→ −∞

as λ → ∞ .

But then, for ﬁxed u = 0 and suﬃciently large λ > 0 we may let u1 = λu, and from Theorem 6.1 we obtain the existence of a non-trivial solution u to (6.1), (6.2). In order to obtain solutions u+ ≥ 0 ≥ u− , we may truncate g above or below u = 0, replacing g by 0, if u ≤ 0, to obtain a positive solution, respectively replacing g by 0, if u ≥ 0, in order to obtain a solution u− ≤ 0. Denote the u truncated functions by g ± (x, u), with primitive G± (x, u) = 0 g ± (x, v) dv. Note that (1◦ ), (2◦ ) remain valid for g ± while (3◦ ) will hold for ±u ≥ R0 , almost everywhere in Ω. Moreover, for ±u ≤ 0 all terms in (3◦ ) vanish. Denote 1 E ± (u) = |∇u|2 dx − G± (x, u) dx 2 Ω Ω

112

Chapter II. Minimax Methods

the functional related to g ± . Then as above E ± ∈ C 1 (H01,2 (Ω)) satisﬁes (P.S.), E ± (0) = 0, and condition (2◦ ) of Theorem 6.1 holds. Moreover, choosing a comparison function u > 0, and letting u1 = λu for large positive, respectively negative λ, by (6.4), (6.5), also condition (3◦ ) of Theorem 6.1 is satisﬁed. Our former reasoning then yields a non-trivial critical point u± of E ± , weakly solving the equation −Δu± = g ± (·, u± )

in Ω .

Since g ± (u) = 0 for ±u ≤ 0, by the weak maximum principle ±u± ≥ 0; see Theorem B.5 of Appendix B. Hence the functions u± in fact are weak solutions of the original Equation (6.1). Finally, if g is H¨older continuous, by (2◦ ) and Lemma B.3 of Appendix B and the remarks following it, u± ∈ C 2,α (Ω) for some α > 0. Hence the strong maximum principle gives u+ > 0 > u− , as desired. (For a related problem a similar reasoning was used by Ambrosetti-Lupo [1].) The Symmetric Mountain Pass Lemma More generally, for problems which are invariant under the involution u → −u, we expect the existence of inﬁnitely many solutions, as in the case of problem (I.2.1), (I.2.3); see Theorem 5.8. However, if a problem does not exhibit the particular homogeneity of problem (I.2.1), (I.2.3), in general it cannot be reduced to a variational problem on the unit sphere in Lp , and a global method is needed. Fortunately, there is a “higher-dimensional” version of Theorem 6.1, especially adapted to functionals with a ZZ2 -symmetry, the symmetric mountain pass lemma – again due to Ambrosetti-Rabinowitz [1]: 6.3 Theorem. Suppose E ∈ C 1 (V ) is even, that is E(u) = E(−u), and satisﬁes (P.-S.). Let V + , V − ⊂ V be closed subspaces of V with codim V + ≤ dim V − < ∞ and suppose V = V − + V + . Also suppose there holds (1◦) E(0) = 0 , (2◦) ∃α > 0, ρ > 0 ∀u ∈ V + : u = ρ ⇒ E(u) ≥ α , (3◦) ∃R > 0 ∀u ∈ V − : u ≥ R ⇒ E(u) ≤ 0 . Then for each j, 1 ≤ j ≤ k = dim V − − codim V + the numbers βj = inf sup E (h(u)) h∈Γ u∈Vj

are critical, where Γ = {h ∈ C 0 (V ; V ); h is odd, h(u) = u if u ∈ V − and u ≥ R} , and where V1 ⊂ V2 ⊂ . . . Vk = V − are ﬁxed subspaces of dimension dim Vj = codim V + + j . Moreover, βk ≥ βk−1 ≥ . . . ≥ β1 ≥ α. For the proof of Theorem 6.3 we need a topological lemma.

6. The Mountain Pass Lemma and its Variants

113

6.4 Intersection Lemma. Let V, V + , V − , Γ, Vj , R be as in Theorem 6.3. Then for any ρ > 0, and any h ∈ Γ there holds γ h(Vj ) ∩ Sρ ∩ V + = j , where Sρ denotes the ρ-sphere Sρ = {u ∈ V ; u = ρ}, and γ denotes the Krasnoselskii genus introduced in Section 5.1; in particular h(Vj )∩Sρ ∩V + = ∅. Proof. Denote Sρ+ = Sρ ∩ V + . For any h ∈ Γ the set A = h(Vj ) ∩ Sρ+ is symmetric, compact, and 0 ∈ A. Thus by Proposition 5.4(5◦ ) there exists a neighborhood U of A such that γ(U) = γ(A). Then γ h(Vj ) ∩ Sρ+ ≥ γ h(Vj ) ∩ Sρ ∩ U ≥ γ (h(Vj ) ∩ Sρ ) − γ (h(Vj ) ∩ Sρ \ U ) . Let Z ⊂ V + be a direct complement of V − and denote by π: V → V − the continuous linear projection onto V − associated with the decomposition V = V − ⊕ Z. Then, since U is a neighborhood of h(Vj ) ∩ Sρ+ , we have π (h(Vj ) ∩ Sρ \ U ) 0, and hence γ (h(Vj ) ∩ Sρ \ U ) ≤ dim V − < ∞ . On the other hand, by Proposition 5.4(2◦ ), (4◦ ) γ (h(Vj ) ∩ Sρ ) ≥ γ h−1 (Sρ ) ∩ Vj . But h(0) = 0, h = id on Vj \ BR (0); hence h−1 (Sρ ) ∩ Vj bounds a symmetric neighborhood of the origin in Vj . Thus, by Proposition 5.2 we conclude that γ (h(Vj ) ∩ Sρ ) ≥ dim Vj = dim V − + j , which implies the lemma. Proof of Theorem 6.3. By Lemma 6.4 we have βj ≥ α for all j ∈ {1, . . . , k}. Moreover, clearly βj+1 ≥ βj for all j. Suppose by contradiction that βj is regular for some j ∈ {1, . . . , k}. For β = βj , N = ∅, and ε = α let ε > 0 and Φ be as constructed in Theorem 3.4. By Remark 3.5(4◦ ) we may assume that Φ(·, t) is odd for any t ≥ 0. Note that by choice of ε we have Φ( ·, t)◦h ∈ Γ for any h ∈ Γ . Choose h ∈ Γ such that there holds E (h(u)) < β + ε for all u ∈ Vj . Then h1 = Φ(·, 1) ◦ h ∈ Γ , and for all u ∈ Vj there holds E (h1 (u)) = E (Φ (h(u), 1)) < β − ε , contradicting the deﬁnition of β. Note that in contrast to Theorem 5.7, our derivation of Theorem 6.3 does not allow us to obtain optimal multiplicity results in the case of degenerate critical values βj = βj+1 = . . . = βk . However, for most applications this defect does not matter. (Actually, by using a notion of pseudo-index, Bartolo-BenciFortunato [1; Theorem 2.4] have been able to prove optimal multiplicity results also in the case of degenerate critical values.) The following result is typical.

114

Chapter II. Minimax Methods

6.5 Theorem. Suppose V is an inﬁnite dimensional Banach space and suppose E ∈ C 1 (V ) satisﬁes (P.-S.), E(u) = E(−u) for all u, and E(0) = 0. Suppose V = V − ⊕ V + , where V − is ﬁnite dimensional, and assume the following conditions: (1◦) ∃α > 0, ρ > 0 : u = ρ, u ∈ V + ⇒ E(u) ≥ α . (2◦) For any ﬁnite dimensional subspace W ⊂ V there is R = R(W ) such that E(u) ≤ 0 for u ∈ W, u ≥ R. Then E possesses an unbounded sequence of critical values. Proof. Choose a basis {ϕ1 , . . .} for V + and for k ∈ IN let Wk = V − ⊕ span {ϕ1 , . . . , ϕk }, with Rk = R(Wk ). Since Wk ⊂ Wk+1 , we may assume that Rk ≤ Rk+1 for all k. Deﬁne classes Γk = h ∈ C 0 (V ; V ) ; h is odd, ∀j ≤ k, u ∈ Wj : u ≥ Rj ⇒ h(u) = u and let βk = inf

sup E (h(u)) .

h∈Γk u∈Wk

Then, by the argument of Theorem 6.3, each of these numbers deﬁnes a critical value βk ≥ α of E. Indeed, for each k we let Wk take the role of V − in Theorem 6.3 and consider the case j = k with Vk = Wk . Observe that for the pseudogradient ﬂow Φ constructed in the proof of Theorem 6.3 there holds Φ(·, t) ∈ Γk for all t. Moreover, Γk ⊂ Γ = h ∈ C 0 (V ; V ); h is odd, h(u) = u if u ∈ Wk and u ≥ Rk . Hence also Lemma 6.4 remains true with Γk instead of Γ , and βk is critical, as claimed. The proof therefore will be complete when we show the following: Assertion. The sequence (βk ) is unbounded. First observe that, since Γk ⊃ Γk+1 while Wk ⊂ Wk+1 for all k, the sequence (βk ) is non-decreasing. Suppose by contradiction that supk βk = β < ∞. Note that K = Kβ is compact and symmetric. Moreover, since β ≥ α > 0 we also have 0 ∈ K. By Proposition 5.4(5◦ ) then γ(K) < ∞, and there exists a symmetric neighborhood N of K such that γ(N ) = γ(Kβ ). Choose ε = α > 0, and let ε ∈]0, ε[ and Φ be determined according to Theorem 3.4 corresponding to ε, β, and N . Observe that Φ may be chosen to be odd, see Remark 3.5(4◦ ), and that by choice of ε for any j ∈ IN we have Φ(u, t) = u for all t ≥ 0 and any u ∈ Wj such that u ≥ Rj ; that is, Φ(·, t) ∈ Γk for any k.

6. The Mountain Pass Lemma and its Variants

115

Let β = β (0) , ε = ε(0) , Φ = Φβ (0) . We iterate the above procedure. For each number β ∈ [α, β − ε] also the set K = Kβ is compact, has ﬁnite genus γ(Kβ ), and possesses a neighborhood N = Nβ with γ(N ) = γ(K ). Moreover, for each such β , with ε = α and N = Nβ we may let εβ ∈]0, ε[, Φβ be determined according to Theorem 3.4. Finitely many neighborhoods ]β (l) − ε(l) , β (l) + ε(l) [, l = 1, . . . , L, where β (l) ∈ [α, β − ε], ε(l) = εβ (l) , cover [α, β − ε]. Clearly, we may assume β = β (0) ≥ β (1) ≥ . . . ≥ β (l) ≥ β (l+1) ≥ . . . ≥ β (L) ≥ α and also that β (l−1) > β (l) + ε(l) > β (l−1) − ε(l−1) > β (l) for all l, 1 ≤ l ≤ L. For any k ∈ IN, with ε = ε(0) > 0 as above, now choose h ∈ Γk satisfying sup E (h(u)) < βk + ε ≤ β + ε . u∈Wk

Composing h with Φ = Φβ (0) yields a map Φ(·, 1) ◦ h =: h ∈ Γk with h (Wk ) ⊂ Eβ−ε ∪ N . Consider the compositions H (m,m) = id, H (m,m−1) = Φβ (m) (·, 1) , H (m,l) = Φβ (m) (·, 1) ◦ . . . ◦ Φβ (l+1) (·, 1),

0≤l <m−1
and let h(0) = Φβ (0) (·, 1)◦h = h , h(l) = Φβ (l) (·, 1)◦h(l−1) = H (l,0) ◦h(0) , for l = 1, . . . , L. By Theorem 3.4, each H (m,l) is the composition of homeomorphisms, hence a homeomorphism itself. By induction, letting N (0) = N, N (l) = Nβ (l) , l = 1, . . . , L, we have h(m) (Wk ) ⊂ Eβ (m) −ε(m) H (m,l) N (l) 0≤l≤m

for all m, 0 ≤ m ≤ L. Moreover, h(m) ⊂ Γk for all m, 0 ≤ m ≤ L. In particular, h(L) ∈ Γk and h(L) (Wk ) ⊂ Eα H (L,l) N (l) . 0≤l≤L

Let Sρ+ = ∂Bρ (0; V + ). Then by assumption (1◦ ) we have H (L,l) N (l) ∩ Sρ+ ⊂ H (L,l) N (l) . h(L) (Wk ) ∩ Sρ+ ⊂ 0≤l≤L

0≤l≤L

By monotonicity and sub-additivity of the genus γ, see Proposition 5.4(2◦ ), (3◦ ), and since each map H (L,l) is a homeomorphism, this implies that

116

Chapter II. Minimax Methods

γ H (L,l) N (l) γ h(L) (Wk ) ∩ Sρ+ ≤ (6.6) =

0≤l≤L

γ (N (l) = γ Kβ (l) =: k0 < ∞ ,

0≤l≤L

0≤l≤L

with k0 independent of k. On the other hand, by the Intersection Lemma 6.4, for any k ∈ IN we have (6.7) γ h(Wk ) ∩ Sρ+ ≥ k . This holds for any h ∈ Γk , in particular, for h = h(L) . Since k is arbitrary, this contradicts (6.6); hence the proof of Theorem 6.5 is complete. Application to Semilinear Equations with Symmetry As an application we state the following existence result for problems of the type (6.1), (6.2) involving odd nonlinearities g. Results of this kind are well known; see the references in Section 5. Note that the symmetry assumption allows removal of any conditions on g near u = 0. 6.6 Theorem. Let Ω be a smoothly bounded domain in IRn , n ≥ 3, and let u g: Ω×IR → IR be a Carath´eodory function with primitive G(·, u) = 0 g(·, v) dv. Suppose: (1◦) g is odd: g(x, −u) = −g(x, u), that is and conditions (2◦ ),(3◦ ) of Theorem 6.2 are satisﬁed, 2n , C : |g(x, u)| ≤ C 1 + |u|p−1 almost everywhere, (2◦) ∃p < 2∗ = n−2 (3◦) ∃q > 2, R0 : 0 < q G(x, u) ≤ g(x, u)u for almost every x, |u| ≥ R0 . Then problem (6.1), (6.2) admits an unbounded sequence (uk ) of solutions uk ∈ H01,2 (Ω). Proof. As in the proof of Theorem 6.2, deﬁne 1 E(u) = |∇u|2 dx − G(x, u) dx . 2 Ω Ω Hypothesis (2◦ ) implies that E is Fr´echet-diﬀerentiable on H01,2 (Ω). A computation similar to (6.3) shows that the assertion of the theorem is equivalent to the assertion that E admits an unbounded sequence of critical values. To prove the latter we invoke Theorem 6.5. Note that by the proof of Theorem 6.2 the functional E satisﬁes (P.-S.); see (6.3). Moreover, since g is odd, E is even: E(u) = E(−u). Finally, E(0) = 0. Denote 0 < λ1 < λ2 ≤ λ3 ≤ . . . the eigenvalues of −Δ on Ω with homogeneous Dirichlet data, as usual, and let ϕj be the corresponding eigenfunctions. We claim that for k0 suﬃciently large there exist ρ > 0, α > 0 such that for all u ∈ V + := span{ϕk ; k ≥ k0 } with uH 1,2 = ρ there holds 0

6. The Mountain Pass Lemma and its Variants

117

∗

E(u) ≥ α. Indeed, by (2◦ ), Sobolev’s embedding H01,2 (Ω) → L2 (Ω), and H¨ older’s inequality, for u ∈ V + we have (with constants C1 , C2 independent of u) 1 |∇u|2 dx − C |u|p dx − C E(u) ≥ 2 Ω Ω 1 ≥ u2H 1,2 − C urL2 up−r (6.8) L2 ∗ − C 0 2 1 −r/2 − C1 λk0 up−2 ≥ u2H 1,2 − C2 H01,2 0 2 where r2 + p−r = 1. In particular, r = n(1 − 2∗ 2 (C2 + 1) and choose k0 ∈ IN such that −r/2 p−2

C1 λk0

ρ

≤

p ) 2∗

> 0, and we may let ρ =

1 4

to achieve E(u) ≥ 1 =: α , for all u ∈ V + with uH 1,2 = ρ . 0

Now ﬁx V + as above and denote V − = span{ϕj ; j < k0 } its orthogonal complement. Finally, on any ﬁnite dimensional subspace W ⊂ H01,2 , by (3◦ ) and (6.4), (6.5) there exist constants Ci = Ci (W ) > 0 such that (6.9)

E(u) ≤ C1 R2 − C2 Rq + C3 → −∞

sup u∈∂BR (0;W )

as R → ∞. Hence, Theorem 6.5 guarantees the existence of an unbounded sequence of critical values βk = inf sup E (h(u)) , k ≥ k0 , h∈Γk u∈Wk

where Wk = span{ϕj ; j ≤ k} and Γk is deﬁned as in the proof of Theorem 6.5. The proof is complete. 6.7 Remark. We can estimate the rate at which the sequence (βk ) diverges. Indeed, by the Intersection Lemma 6.4 , letting Vk = span{ϕj ; j ≥ k}, Wk = span{ϕ1 . . . , ϕk } as in the proof above, we have h(Wk ) ∩ Sρ ∩ Vk = ∅ for any ρ > 0, any k ∈ IN. Hence, by (6.8), with r = n 1 − we obtain

p 2∗

= p−

n(p−2) 2

118

Chapter II. Minimax Methods

βk ≥ sup

inf

ρ>0 u∈Sρ ∩Vk

≥ sup ρ>0

E(u)

1 −r/2 p−2 ρ2 − C2 − C1 λk ρ 2 p

r

n

≥ cλk p−2 = cλk ( p−2 − 2 ) , for k large. Finally, by the asymptotic formula (see Weyl [1], or EdmundsMoscatelli [1] ) λk k −2/n → c > 0 (k → ∞) , and it follows that with a constant c > 0 we have 2p

βk ≥ ck n(p−2) −1

(6.10)

for large k. Estimate (6.10) will prove useful in the next section.

7. Perturbation Theory A natural question, which even today is not adequately settled, is whether the symmetry of the functional is important for results like Theorem 6.5 to hold. A partial answer for problems of the kind studied in Theorem 6.6 was independently obtained by Bahri-Berestycki [1] and Struwe [1] in 1980. In abstract form, the variational principle underlying these results later ˜ for the nonwas phrased by Rabinowitz [10]. It will be convenient to write E symmetric functional and to reserve the notation E for its even symmetrization. ˜ ∈ C 1 (V ) satisﬁes (P.-S.). Let W ⊂ V be a ﬁnite 7.1 Theorem. Suppose E dimensional subspace of V , w∗ ∈ V \ W , and let W ∗ = W ⊕ span{w∗ }; also let W+∗ = {w + tw∗ ; w ∈ W, t ≥ 0} denote the “upper half-space” in W ∗ . Suppose ˜ =0, (1◦) E(0) ◦ ˜ ≤0, (2 ) ∃R > 0 ∀u ∈ W : u ≥ R ⇒ E(u) ◦ ∗ ∗ ∗ ˜ ≤0, (3 ) ∃R ≥ R ∀u ∈ W : u ≥ R ⇒ E(u) and let ˜ ˜ Γ = {h ∈C 0 (V, V ) ; h is odd, h(u) = u if max{E(u), E(−u)} ≤0, ∗ in particular, if u ∈ W, u ≥ R, or if u ∈ W , u ≥ R∗ } . Then, if β ∗ = inf

˜ ∗ (u)) > β = inf sup E ˜ (h(u)) ≥ 0 , sup E(h

h∈Γ u∈W ∗ +

h∈Γ u∈W

˜ possesses a critical value ≥ β ∗ . the functional E Proof. For γ ∈]β, β ∗ [ let

7. Perturbation Theory

119

˜ (h(u)) ≤ γ for u ∈ W } . Λ = {h ∈ Γ ; E By deﬁnition of β, clearly Λ = ∅. Hence γ ∗ = inf

˜ (h(u)) ≥ β ∗ sup E

h∈Λ u∈W ∗ +

is well-deﬁned. We contend that γ ∗ is critical. Assume by contradiction that γ ∗ is regular and choose ε > 0, Φ according to Theorem 3.4, corresponding to γ ∗ , ε = γ ∗ − γ > 0, N = ∅. Also select h ∈ Λ with ˜ (h(u)) < γ ∗ + ε . sup E ∗ u∈W+

Deﬁne an odd map h : W ∗ → V

Φ (h(u), 1) , if u ∈ W+∗ h (u) = −Φ (h(−u), 1) , if −u ∈ W+∗ . Note that by choice of ε and since h ∈ Λ we have Φ (h(−u), 1) = h(−u) = −h(u) = −Φ (h(u), 1) for u ∈ W . Hence h is well-deﬁned, odd, and continuous. Moreover, since ˜ ≤0 0 ≤ β < γ < γ ∗ − ε, the map Φ(·, 1) ﬁxes any point u that satisﬁes E(u) ˜ and E(−u) ≤ 0. By deﬁnition of Λ so does h, and hence the composition Φ(·, 1) ◦ h. It follows that h may be extended to a map h ∈ Λ. But now the estimate ˜ (h (u)) = sup E ˜ (Φ (h(u), 1)) < γ ∗ − ε sup E ∗ u∈W+

∗ u∈W+

yields the desired contradiction. ˜ satisfying the Theorem 7.1 suggests comparing a non-symmetric functional E ◦ ◦ remaining hypotheses (1 ), (2 ) of Theorem 6.5 with its symmetrization 1˜ ˜ E(u) + E(−u) . E(u) = 2 Then Theorem 6.5 applies to E and we obtain an unbounded sequence (βk ) of critical values of E. Playing the speed of divergence βk → ∞ oﬀ against ˜ from Theorem 7.1 we can glean the the perturbation from symmetry E − E, existence of inﬁnitely many critical points for functionals perturbed from symmetry. Stating a general theorem of this kind, however, seems to involve so many technical conditions that any reader would ask himself if these conditions can ever be met in practice. Therefore we prefer to present such an application immediately.

120

Chapter II. Minimax Methods

Applications to Semilinear Elliptic Equations We return to the setting of problem (6.1), (6.2) under (essentially) the assumptions (1◦ )–(3◦ ) of Theorem 6.6 studied previously. However, in order to make estimate (6.4) uniform in x ∈ Ω, we assume that g is continuous in all its variables. For convenience we restate also the remaining conditions. 7.2 Theorem. Let Ω be a smoothly bounded domain in IRn , n ≥ 3.

uSuppose with primitive G(x, u) = g(x, v) dv; (1◦) g: Ω×IR → IR is continuous and odd 0 2n , C : |g(x, u)| ≤ C 1 + |u|p−1 almost everywhere; (2◦) ∃p < 2∗ = n−2 (3◦) ∃q > 2, R0 : 0 < q G(x, u) ≤ g(x, u)u for almost every x, |u| ≥ R0 . Moreover, suppose that q 2p −1> . n(p − 2) q−1 Then for any f ∈ L2 (Ω) the problem −Δu = g(·, u) + f u=0

in Ω , on ∂Ω ,

has an unbounded sequence of solutions uk ∈ H01,2 (Ω), k ∈ IN. Proof. The assertion of the theorem is equivalent to the claim that the functional 1 ˜ |∇u|2 dx − G(·, u) dx − f u dx E(u) = 2 Ω Ω Ω on H01,2 (Ω) possesses an unbounded sequence of critical values. ˜ Clearly, E(0) = 0, and on any ﬁnite dimensional subspace W ⊂ H01,2 (Ω) ◦ by (3 ) and (6.4), (6.5) the estimate (7.1)

sup

˜ E(u) ≤ C1 ρ2 − C2 ρq + C3 ρ + C4 → −∞

(ρ → ∞)

u∈∂Bρ (0;W )

holds with constants Ci = Ci (W ) > 0. (The term Ω f u dx was estimated by using H¨ older’s inequality.) To see (P.-S.), as in (6.3) in the proof of Theorem 6.2, for a (P.-S.) sequence (um ) consider the estimate ˜ m ) − um , DE(u ˜ m ) C + o(1) um H 1,2 ≥ q E(u 0

= q E(um ) − um , DE(um ) − (q − 1) ≥

q−2 um 2H 1,2 − Cum H 1,2 − C , 0 0 2

f u dx Ω

from which boundedness of (um ) in H01,2 (Ω) follows at once. (Note that the ˜ ˜ + E(−u) equals the functional E symmetrized functional E(u) = 12 E(u) considered in Theorem 6.6.) Using Proposition 2.2, the existence of a strongly

7. Perturbation Theory

121

convergent subsequence of (um ) now is established exactly as in the proof of Theorem 6.2. Let λj , ϕj , V + = span{ϕk ; k ≥ k0 }, Wk = span{ϕj ; j ≤ k} as in the proof of Theorem 6.6. By (7.1) we may choose a non-decreasing sequence (Rk ) ˜ such that E(u) ≤ 0 for all u ∈ Wk with uH 1,2 ≥ Rk ; hence also E(u) ≤ 0 for 0 all such u. With this sequence (Rk ) let (Γk ) be deﬁned as in Theorem 6.5 and for k ≥ k0 let 2p

βk = inf sup E (h(u)) ≥ c k n(p−2) −1 − β0

(7.2)

h∈Γk u∈Wk

be the sequence of critical values of the symmetrized functional E constructed ˜ does in Theorem 6.6, the lower estimate following from (6.10). Suppose that E not have any critical values larger than some number β − 1. Fix k ≥ k0 such that βk ≥ β for k ≥ k. We will use Theorem 7.1 to derive a uniform bound βk ≤ C,

for all k ≥ k ,

which will yield the desired contradiction to (7.2). For k ≥ k let W = Wk , w∗ = ϕk+1 , W ∗ = Wk+1 , and let Γ be deﬁned as in Theorem 7.1, with R = Rk , R∗ = Rk+1 . Note that Γ ⊂ Γk+1 ⊂ Γk . Since ˜ from Theorem 7.1 by assumption there are no critical values β > β − 1 for E, we conclude that β˜∗ := β˜k∗ = inf ∗

˜ (h∗ (u)) sup E

h ∈Γ u∈W ∗ +

˜ (h(u)) . = β˜ := β˜k = inf sup E h∈Γ u∈W

From oddness of any map h ∈ Γ and the estimate 1˜ ˜ ˜ ˜ E(u) = E(u) + E(−u) ≤ max{E(u), E(−u)} , 2 valid for all u, we also deduce that (7.3) β˜k ≥ inf sup E h(u) ≥ βk ≥ β , for all k ≥ k . h∈Γ u∈W

Given ε > 0 (say ε = 1) choose a map h∗ ⊂ Γ such that ˜ h∗ (u) ≤ sup E ˜ h∗ (u) < β˜∗ + ε = β˜k + ε . sup E k u∈W

∗ u∈W+

Let · denote the norm in H01,2 (Ω), respectively H −1 (Ω). Consider a pseudogradient vector ﬁeld v for E such that v ≤ 2, v(u), DE(u) ≥ min 1, DE(u) DE(u) . Since E is even, we may assume v to be odd. Note that

122

Chapter II. Minimax Methods

˜ ˜ v(u), DE(u) = v(u), DE(u) + v(u), DE(u) − DE(u) ˜ ≥ min{1, DE(u)} DE(u) − 2DE(u) − DE(u) ˜ ˜ ˜ ≥ min{1, DE(u)} DE(u) − 4 DE(u) − DE(u) . ˜ oﬀ N ˜δ for any δ ≥ 8, That is, v also is a pseudo-gradient vector ﬁeld for E ˜δ is given by where for δ > 0 the set N ˜ ˜ ˜ ˜δ = u ∈ H 1,2 (Ω) ; DE(u) − DE(u) > δ −1 min 1, DE(u) DE(u) . N 0 ˜δ is a neighborhood of the set Note that – unless f ≡ 0, which is trivial – N ˜δ ∪ −N ˜δ the ˜ of critical points of E and E, for any δ > 1. Denote Nδ = N ˜δ . Let 0 ≤ η ≤ 1 be an even, locally Lipschitz function symmetrized sets N satisfying η(u) = 0 in N10 and η(u) = 1 oﬀ N = N20 . Also let ϕ be Lipschitz, 0 ≤ ϕ ≤ 1, ϕ(s) = 0 for s ≤ 0, ϕ(s) = 1 for s ≥ 1, and let ˜ ˜ e(u) = −ϕ max{E(u), E(−u)} η(u)v(u) denote the truncated pseudo-gradient vector ﬁeld on H01,2 (Ω). Then, if Φ denotes the (odd) pseudo-gradient ﬂow for E induced by e, Φ ˜ and will strictly decrease E ˜ oﬀ N with will also be a pseudo-gradient ﬂow for E “speed” 1 d ˜ 2 ˜ E Φ(u, t) |t=0 ≤ − min 1, DE(u) , dt 2

˜ if E(u) ≥1.

Moreover, Φ(·, t) ∈ Γ for all t ≥ 0. ˜ achieves its supremum on Φ h∗ (W ∗ ), t Composing h∗ with Φ(·, t), unless E at a point u ∈ N , thus we obtain that d ˜ Φ h∗ (u), t ≤ −c < 0 , sup E dt u∈W ∗ uniformly in t ≥ 0. (Here, condition (P.-S.) was used; the constant c may depend on our initial choice for the map h.) Replacing h∗ by Φ(·, t) ◦ h∗ and ˜ achieves its suprechoosing t large, if necessary, we hence may assume that E mum on h∗ (W ∗ ) in N . That is, we may estimate ˜ (h∗ (u)) − β˜k + ε = β˜k∗ + ε ≥ sup E u∈W ∗

(7.4)

˜ (h∗ (u)) − = sup E u∈W ∗

≥ β˜k+1 − c

sup u∈N ˜ ˜ +ε E(u)≤ β k

sup u∈N, ˜ ˜ +ε E(u)≤ β k

sup u∈N, ˜ ˜ +ε E(u)≤ β k

˜ ˜ |E(u) − E(−u)| |2

f u dx| Ω

uL2 .

˜ But for u ∈ N with E(u) ≤ β˜k + ε, with constants c > 0 from (3◦ ) and (6.4), (6.5) we obtain

7. Perturbation Theory

q−2 uqL2 ≤ c uqLq ≤ c G(u) dx + C 2 Ω 1 g(u)u − G(u) dx + C ≤c 2 Ω 1 1 ˜ ˜ − u, DE(u) + f u dx + C = c E(u) 2 2 Ω ˜ ˜ ≤ c E(u) + c 1 + DE(u) uH 1,2 + C 0 ˜ ˜ ≤ c E(u) + c 1 + DE(u) − DE(u) uH 1,2 + C 0

≤ c β˜k + c uH 1,2 + C . 0

From this there follows

u2H 1,2 0

˜ ≤ 2 E(u) +2

G(u) dx + CuL2 Ω

≤ c β˜k + c uH 1,2 + C . 0

Together with the former estimate this shows that for such u there holds 1/q 1/q uL2 ≤ c β˜k + C ≤ c β˜k .

Inserting this estimate in (7.4) above yields the inequality 1−q 1/q q ˜ ˜ ˜ ˜ ˜ , βk+1 ≤ βk + c βk = βk 1 + c βk for all k ≥ k0 , with a uniform constant c. By iteration therefore β˜k0 +l ≤ β˜k0

k04 +l−1

1−q

1 + c β˜k q

k=k0

k

0 +l−1

≤ β˜k0 exp

1−q q ln 1 + c β˜k

k=k0

≤ β˜k0 exp c

k0 +l−1

1−q q

β˜k

.

k=k0

But by (7.2), (7.3) we have 2p

β˜k ≥ βk ≥ c k n(p−2) −1 . Since we assume that μ=

1−q · q

we can uniformly estimate

2p −1 n(p − 2)

< −1 ,

123

124

Chapter II. Minimax Methods

β˜k0 +l ≤ β˜k0 exp c

k0 +l−1

1−q q

β˜k

k=k0

∞ μ ˜ k <∞, ≤ βk0 exp c k=k0

˜ possesses an for all l ∈ IN, which yields the desired contradiction. Thus, E unbounded sequence of critical values and the proof is complete. Remark 7.3. (1◦ ) In the special case g(x, u) = u|u|p−2 , by Theorem 7.2 for any f ∈ L2 the problem (7.5) (7.6)

−Δu = u|u|p−2 + f u=0

in Ω , on ∂Ω ,

admits inﬁnitely many solutions, provided 2 < p < p∗ , where p∗ > 2 is the largest root of the equation (7.7)

1 2p∗ =2+ ∗ . n(p∗ − 2) p −1

Observe that p∗ < 2∗ . (2◦) Recently, Bahri-Lions [1] have been able to improve the estimate for βk as follows 2p βk ≥ c k n(p−2) , which yields the improved bounds q 2p < q−1 n(p − 2) for p and q for which Theorem 7.2 is valid. In particular, for the model problem (7.5), (7.6) above, one obtains inﬁnitely many solutions, provided 2 < p < p¯, where 2 < p¯ < 2∗ solves the equation 1 2¯ p =1+ . n(¯ p − 2) p¯ − 1 (3◦) Bahri [1] has shown that for any p ∈]2, 2∗ [ there is a dense open set of f ∈ H −1 (Ω) for which problem (7.5), (7.6) possesses inﬁnitely many solutions. Generally, it is expected that this should be true for all f (for instance in L2 (Ω)), also for more general problems of type (6.1), (6.2) considered earlier. Indeed, this is true in the case of ordinary diﬀerential equtions – see for example Nehari [1], Struwe [2], or Turner [1] – and in the radially symmetric case, where diﬀerent solutions of (6.1), (6.2) can be characterized by the nodal properties they possess; see Struwe [3], [4]. However, the proof of this conjecture in general remains open. (4◦) Applications of perturbation theory to Hamiltonian systems are given by Bahri-Berestycki [2], Rabinowitz [10], and – more recently – by Long [1]. (5◦) Finally, Tanaka [1] has obtained similar perturbation results for nonlinear wave equations, as in Section I.6.

8. Linking

125

8. Linking In the preceding chapter we have seen that symmetry is not essential for a variational problem to have “many” solutions. In this chapter we will describe another method for dealing with non-symmetric functionals. This method was introduced by Benci [1], Ni [1], and Rabinowitz [8]. Subsequently, it was generalized to a setting possibly involving also “indeﬁnite” functionals by BenciRabinowitz [1]. It is based on the topological notion of “linking”. 8.1 Definition. Let S be a closed subset of a Banach space V , Q a submanifold of V with relative boundary ∂Q. We say S and ∂Q link if (1◦) S ∩ ∂Q = ∅, and (2◦) for any map h ∈ C 0 (V ; V ) such that h|∂Q = id there holds h(Q) ∩ S = ∅. More generally, if S and Q are as above and if Γ is a subset of C 0 (V ; V ), then S and ∂Q will be said to link with respect to Γ , if (1◦ ) holds and if (2◦ ) is satisﬁed for any h ∈ Γ . 8.2 Example. Let V = V1 ⊕ V2 be decomposed into closed subspaces V1 , V2 , where dim V2 < ∞. Let S = V1 , Q = BR (0; V2 ) with relative boundary ∂Q = {u ∈ V2 ; u = R}. Then S and ∂Q link.

Fig. 8.1.

Proof. Let π: V → V2 be the (continuous) projection of V onto V2 , and let h be any continuous map such that h|∂Q = id. We have to show that π h(Q) 0. For t ∈ [0, 1], u ∈ V2 deﬁne ht (u) = t π h(u) + (1 − t)u .

126

Chapter II. Minimax Methods

Note that ht ∈ C 0 (V2 ; V2 ) deﬁnes a homotopy of h0 = id with h1 = π ◦ h. Moreover, ht |∂Q = id for all t. Hence the topological degree deg(ht , Q, 0) is well-deﬁned for all t. By homotopy invariance and normalization of the degree (see for instance Deimling [1; Theorem 1.3.1]), we have deg(π ◦ h, Q, 0) = deg(id, Q, 0) = 1 . Hence 0 ∈ π ◦ h(Q), as was to be shown. 8.3 Example. Let V = V1 ⊕ V2 be decomposed into closed subspaces with dim V2 < ∞, and let u ∈ V1 with u = 1 be given. Suppose 0 < ρ < R1 , 0 < R2 and let S = {u ∈ V1 ; u = ρ} , Q = {su + u2 ; 0 ≤ s ≤ R1 , u2 ∈ V2 , u2 ≤ R2 } , with relative boundary ∂Q = {su + u2 ∈ Q ; s ∈ {0, R1 } or u2 = R2 }. Then S and ∂Q link.

Fig. 8.2.

Proof. Let π: V → V2 denote the projection onto V2 , and let h ∈ C 0 (V ; V ) show that there exists u ∈ Q such that the relations satisfy h|∂Q = id.We must h(u) = ρ and π h(u) = 0 simultaneously hold. For t ∈ [0, 1], s ∈ IR, u2 ∈ V2 let ht (s, u2 ) = th(u) − π h(u) + (1 − t)s − ρ, tπ h(u) + (1 − t)u2 , where u = su + u2 . This deﬁnes a family of maps ht : IR × V2 → IR × V2 depending continuously on t ∈ [0, 1]. Moreover, if u = su + u2 ∈ ∂Q, we have ht (s, u2 ) = tu − u2 + (1 − t)s − ρ, u2 = (s − ρ, u2 ) = 0

8. Linking

127

for all t ∈ [0, 1]. Hence, if we identify Q with a subset of IR × V2 via the decomposition u = su + u2 , the topological degree deg(ht , Q, 0) is well-deﬁned and by homotopy invariance deg(h1 , Q, 0) = deg(h0 , Q, 0) = 1 , where h0 (s, u2 ) = (s − ρ, u2 ). Thus, there exists u = su + u2 ∈ Q such that h1 (u) = 0, which is equivalent to π h(u) = 0

and

h(u) = ρ ,

as desired. 8.4 Theorem. Suppose E ∈ C 1 (V ) satisﬁes (P.-S.). Consider a closed subset S ⊂ V and a submanifold Q ⊂ V with relative boundary ∂Q. Suppose (1◦) S and ∂Q link, (2◦) α = inf u∈S E(u) > supu∈∂Q E(u) = α0 . Let Γ = {h ∈ C 0 (V ; V ) ; h|∂Q = id} . Then the number

β = inf sup E h(u) h∈Γ u∈Q

deﬁnes a critical value β ≥ α of E. Proof. Suppose by contradiction that Kβ = ∅. For ε = α − α0 > 0, N = ∅ let ε > 0 and Φ: V × [0, 1] → V be the pseudo-gradient ﬂow constructed in Theorem 3.4. Notethat by choice of ε there holds Φ(·, t)|∂Q = id for all t. Let h ∈ Γ such that E h(u) < β + ε for all u ∈ Q. Deﬁne h = Φ(·, t) ◦ h. Then h ∈ Γ and sup E h(u) < β − ε u∈Q

by Theorem 3.4(3◦ ), contradicting the deﬁnition of β. Remark. By Example 8.3, letting S = ∂Bρ (0; V ), Q = {tu ; 0 ≤ t ≤ 1} with relative boundary ∂Q = {0, u}, 0 < ρ < u, Theorem 8.4 contains the mountain pass lemma Theorem 6.1 as a special case. In contrast to Theorem 6.1, the above Theorem 8.4 also allows higher-dimensional comparison sets Q, similar to Theorem 6.3 in the symmetric case.

128

Chapter II. Minimax Methods

Applications to Semilinear Elliptic Equations We demonstrate this with yet another variant of Theorem 6.2, allowing for linear growth of g near u = 0. Conceivably, condition (1◦ ) below may be further weakened. However, we will not pursue this. 8.5 Theorem. Let Ω be a smooth, bounded domain in IRn , n ≥ 3, and let g: Ω × IR → IR be measurable in x ∈ Ω, diﬀerentiable in u ∈ IR with derivative gu , and with primitive G. Moreover, suppose that (1◦) g(x, 0) = 0, and g(x, u) ≥ gu (x, 0), for almost every x ∈ Ω, u ∈ IR ; u 2n (2◦) ∃p < n−2 , C : gu (x, u) ≤ C 1 + |u|p−2 , for almost every x ∈ Ω, u ∈ IR; (3◦) ∃q > 2, R0 : 0 < q G(x, u) ≤ g(x, u)u , for almost every x ∈ Ω, if |u| ≥ R0 . Then the problem (8.1)

−Δu = g(·, u)

(8.2)

u=0

in Ω on ∂Ω

admits a solution u ≡ 0. Proof. For u ∈ H01,2 (Ω) let E(u) =

1 2

|∇u|2 dx −

Ω

G(x, u) dx . Ω

By assumptions (2◦ ), (3◦ ) the functional E is of class C 2 and as in the proof of Theorem 6.2 satisﬁes (P.-S.) on H01,2 (Ω). It suﬃces to show that E admits a critical point u ≡ 0. Note that by assumption (1◦ ) problem (8.1), (8.2) admits u ≡ 0 as a trivial solution. Let ϕk denote the eigenfunctions of the linearized equation −Δϕk = gu (x, 0)ϕk + λk ϕk ϕk = 0

in Ω on ∂Ω

with eigenvalues λ1 < λ2 ≤ λ3 ≤ . . . . Denote k0 = min{k ; λk > 0} and let V + = span{ϕk ; k ≥ k0 }, V − = span{ϕ1 , . . . , ϕk0 −1 }. Note that by (1◦ ) we have u 1 G(x, u) = g(x, v) dv ≥ gu (x, 0)u2 . 2 0 Hence there holds (8.3)

E(u) ≤

1 2 D E(0)(u, u) ≤ 0, for u ∈ V − , 2

while by deﬁnition of V + clearly we have

8. Linking

129

D2 E(0)(u, u) ≥ λko u2L2 , for u ∈ V + . To strengthen the latter inequality note that by (2◦ ) the function gu (x, 0) is essentially bounded. Thus for suﬃciently large k1 we have

|∇u|2 dx −

D2 E(0)(u, u) = 0

gu (x, 0)u2 dx ≥ Ω

1 u2H 1,2 0 2

uniformly for u ∈ span{ϕk ; k ≥ k1 }. Since the complement of this space in V + has ﬁnite dimension, we conclude that there exists λ > 0 such that D2 E(0)(u, u) ≥ λu2H 1,2 , 0

uniformly for u ∈ V + . But E ∈ C 2 H01,2 (Ω) ; it follows that for suﬃciently small ρ > 0 we have (8.4)

inf E(u) ≥

u∈Sρ+

λρ2 1 2 D E(0)(u, u) − o u2H 1,2 ≥ >0, 0 2 4

where Sρ+ = {u ∈ V + ; uH 1,2 = ρ} and where o(s)/s → 0 (s → 0) . 0 By (3◦ ) ﬁnally, as in the proof of Theorem 6.2 (see (6.4), (6.5)), for any ﬁnite dimensional subspace W we have E(u) → −∞,

as u → ∞, u ∈ W .

Recalling (8.3), (8.4), we see that the assumptions of Theorem 8.4 are satisﬁed with S = Sρ+ and Q = {u− + sϕk0 ; u− ∈ V − , u− H 1,2 ≤ R, 0 ≤ s ≤ R} , 0

for suﬃciently large R > 0. Thus E admits a critical point u with E(u) ≥ The proof is complete.

λρ2 . 4

Further applications of Theorem 8.4 to semilinear elliptic boundary value problems are given in the survey notes by De Figueireido [1] or Rabinowitz [11; p. 25 ﬀ.]. In particular, Theorem 8.4 oﬀers a simple and uniﬁed approach to asymptotically linear equations, possibly “resonant” at u = 0 or at inﬁnity, as in Ahmad-Lazer-Paul [1], Amann [4], Amann-Zehnder [1]. (See for instance Rabinowitz [11; Theorem 4.12], or Bartolo-Benci-Fortunato [1] and the references cited therein. See also Chang [1; p. 708].)

130

Chapter II. Minimax Methods

Applications to Hamiltonian Systems A more reﬁned application of linking is given in the next theorem due to Hofer and Zehnder [1], prompted by the work of Viterbo [1]. Once again we deal with Hamiltonian systems (8.5)

x˙ = J ∇H(x) ,

where H is a given smooth Hamiltonian and J is the skew-symmetric matrix 0 −id J = id 0 on IR2n = IRn × IRn . 8.6 Theorem. Let H ∈ C 2 (IR2n ; IR). Suppose that 1 is a regular value of H and S = S1 = H −1 (1) is compact and connected. Then for any δ > 0 there is a number β ∈]1 − δ, 1 + δ[ such that Sβ = H −1 (β) carries a periodic solution of the Hamiltonian system (8.5). (Note that by the implicit function theorem and compactness of S there exists a number δ0 > 0 such that for any β ∈]1 − δ0 , 1 + δ0 [ the hypersurface Sβ is of class C 2 , compact, and diﬀeomorphic to S.) Remarks. By Theorem 8.6, in order to obtain a periodic solution of (8.5) on the ﬁxed surface S = S1 , it would suﬃce to obtain a-priori bounds (in terms of the action integral) for periodic solutions to (8.5) on surfaces Sβ near S. Benci-Hofer-Rabinowitz [1] have shown that this is indeed possible, provided certain geometric conditions are satisﬁed, including, for instance, the condition that S be strictly star-shaped with respect to the origin. However, also energy surfaces of Hamiltonians that are only convex either in the position or in the momentum variables are allowed. The existence of periodic solutions to (8.5) on S likewise can be derived if S is of “contact type”. This notion, introduced by Weinstein [3], allows one to give an intrinsic interpretation of the existence results by Rabinowitz [5] and Weinstein [2] for closed trajectories of Hamiltonian systems on convex or strictly star-shaped energy hypersurfaces. See Hofer-Zehnder [1], or Zehnder [2] for details. Proof of Theorem 8.6. Observe that, as remarked earlier in the proof of Theorem I.6.5, the property that a level hypersurface Sβ = H −1 (β) carries a periodic solution of (8.5) is independent of the particular Hamiltonian H having Sβ as a level surface. We now use this freedom to redeﬁne H suitably. For 0 < δ < δ0 let U = H −1 [1 − δ, 1 + δ] # S × [−1, 1]. Then IR2n \ U has two components. Indeed, by Alexander duality

8. Linking

131

˜ 0 (IR2n \ U ; ZZ) # H 2n−1 (U ; ZZ) # H 2n−1 (S ; ZZ) # ZZ , H in Spanier’s [1] notation. Denote by A the unbounded component of IR2n \ U and by B the bounded component. We may assume 0 ∈ B. Also let γ = diam U > 0. Fix numbers r, b > 0 such that γ 0 for −δ < s < δ. Also let g: IR → IR be a smooth function satisfying (8.6)

3 3 2 πs for s > r, g(s) = πs2 for large s , 2 2 and 0 < g (s) ≤ 3πs for s > r . g(s) = b for s ≤ r, g(s) ≥

Then deﬁne ⎧ 0, ⎪ ⎨ f (s), ˜ H(x) = ⎪ b, ⎩ g(|x|),

if if if if

x ∈ B, H(x) = 1 + s, −δ ≤ s ≤ δ, x ∈ A, |x| ≤ r, |x| > r .

Note that for x ∈ IR2n we can estimate (8.7)

3 3 ˜ −b + π|x|2 ≤ H(x) ≤ π|x|2 + b . 2 2

˜ replacing H. Such Now we look for solutions of (8.5) of period 1, with H solutions uniquely correspond to the critical points of the functional 1 1 1 ˜ E(x) = x, ˙ J x dt − H(x) dt 2 0 0 on the space of 1-periodic (C 1 −)maps x: IR → IR2n . By the following lemma we are able to distinguish a periodic solution on a level hypersurface Sβ . ˜ replacing 8.7 Lemma. Suppose x is a 1-periodic C 1 -solution of (8.5), with H H, and which satisﬁes E(x) > 0. Then x(t) ∈ Sβ for some β ∈ [1 − δ, 1 + δ], for all t. ˜ ≥ 0, any constant solution x of (8.5) satisﬁes E(x) ≤ 0. Assume Proof. Since H ˜ x(t0 ) = H ˜ x(t) = xis non-constant with |x(t0 )| > r. Then g |x(t0 )| = H g |x(t)| also for t close to t0 and by (8.6) it follows that |x(t)| ≡ |x(t0 )| = s0 . In particular, x satisﬁes

132

Chapter II. Minimax Methods

x˙ = J and we compute

g (s0 ) x s0

1 ˜ x(t0 ) g (s0 ) s0 − H 2 3 3 ≤ πs20 − πs20 = 0 , 2 2

E(x) =

˜ The claim now follows. by deﬁnition of H. In order to exhibit a 1-periodic solution x of (8.5) with E(x) > 0, a variational argument involving linking will be applied. Denote by V = H 1/2,2 (S 1 ; IR2n ) the Hilbert space of all 1-periodic functions exp (2πktJ ) xk ∈ L2 [0, 1] ; IR2n , (8.8) x(t) = k∈ZZ

with Fourier coeﬃcients xk ∈ IR2n satisfying 2 2 |k| |xk | + |x0 |2 < ∞ . x = 2π k∈ZZ

The inner product on V is given by |k|xk , yk + x0 , y0 (x, y)V := 2π k∈ZZ

inducing the norm · above. Split V = V − ⊕ V 0 ⊕ V + orthogonally, where V − = {x ∈ V ; xk = 0 for k ≥ 0} , V 0 = {x ∈ V ; xk = 0 for k =

0} , V + = {x ∈ V ; xk = 0 for k ≤ 0} with reference to the Fourier decomposition (8.8) of an element x ∈ V . Also denote P − , P 0 and P + the corresponding projections. Thus, any x ∈ V may be uniquely expressed x = x− + x0 + x+ , where x− = P − x, etc. Note that the self-adjoint operator Lx = −J x˙ has eigenspaces Vk = {exp(2πktJ )xk ; xk ∈ IR2n } with eigenvalues 2πk, k ∈ ZZ. Moreover, dim Vk = 2n, for all k. Hence the related quadratic form

8. Linking

1 A(x) = 2

1

0

133

⎧ ⎨ < 0, for x ∈ V − , x, ˙ J x dt = 0, for x ∈ V 0 , ⎩ > 0, for x ∈ V + .

In fact, using the projections P ± , A can be written A(x) =

1 (−P − + P + )x, x V . 2

˜ is of class C 2 and behaves like |x|2 for large |x|, its mean Note that since H G(x) =

1

˜ H(x) dt

0

is of class C 2 on L2 (S 1 ; IR2n ). Hence also E ∈ C 2 (H 1/2,2 (S 1 )). Restricting G to V , from compactness of the embedding H 1/2,2 (S 1 ) → L2 (S 1 ) we also deduce that DG is a completely continuous map from V into its dual. 8.8 Lemma. There exist numbers α > 0, ρ ∈]0, 1[ such that E(x) ≥ α for x ∈ V + , x = ρ. ˜ ˜ = 0 implies that DG(0) = 0, D2 G(0) = 0. Hence Proof. ∇H(0) = 0, ∇2 H(0) D2 E(0) = −P − + P + . Since E ∈ C 2 , E(0) = 0, DE(0) = 0, the lemma follows. Now deﬁne Q = {x = x− + x0 + s e ∈ V ; x− + x0 ≤ R, and 0 ≤ s ≤ R} , with R > 1 to be determined, and with a unit vector 1 e = √ exp (2πtJ )a ∈ V + , 2π

|a| = 1 .

Denote ∂Q the relative boundary of Q in V − ⊕ V 0 ⊕ IR · e; that is, ∂Q = x− + x0 + s e ∈ Q ; x− + x0 = R or s ∈ {0, R} . 8.9 Lemma. If R is suﬃciently large, then E|∂Q ≤ 0. Proof. Note that (8.7) implies that G(x) = 0

1

3 ˜ H(x) dt ≥ −b + πx2L2 . 2

Therefore, for x = x− +x0 +se with s = R or x− +x0 2 = x− 2 +x0 2L2 = R, we obtain

134

Chapter II. Minimax Methods

1 (−P − + P + )x, x V − G(x) 2 1 3 s2 1 ≤ − x− 2 + s2 + b − π x− 2L2 + x0 2L2 + 2 2 2 2π 1 − 2 1 2 3 0 2 ≤ − x − s − πx L2 + b 2 4 2 ≤0,

E(x) =

˜ ≥ 0 implies G ≥ 0 and therefore if R > 0 is suﬃciently large. Moreover, H − 0 E ≤ 0 on V ⊕ V , that is at s = 0, which concludes the proof. Fix 0 < ρ < R as in Lemmas 8.8, 8.9. Denote Sρ+ = {x ∈ V + ; x = ρ}. Deﬁne a class Γ of maps V → V as follows: Γ is the class of maps h ∈ C 0 (V ; V ) such that h is homotopic to the identity through a family of maps ht = Lt + Kt , 0 ≤ t ≤ T , where L0 = id, K0 = 0, and where for each t there holds: + 0 − ⊕V0⊕V+ →V−⊕V0⊕V+ Lt = (L− t , Lt , Lt ): V

is a linear isomorphism preserving sub-spaces, Kt is compact and ht (∂Q)∩Sρ+ = ∅, for each t. Following Benci-Rabinowitz [1], we now establish 8.10 Lemma. ∂Q and Sρ+ link with respect to Γ . Proof. Choose h ∈ Γ . We must show that h(Q) ∩ Sρ+ = ∅, or equivalently that the equations − h(x) = ρ (8.9) (P + P 0 ) ◦ h (x) = 0, are satisﬁed for some x ∈ Q. Using a degree argument as in Example 8.3, we establish (8.9) for every ht in the family deﬁning h ∈ Γ . Consider Q ⊂ V − ⊕ V 0 ⊕ IR · e =: W and to (8.9), these equations become represent x = x− + x0 + se. Applying L−1 t x− + x0 +(P − + P 0 )L−1 t Kt (x) = 0 , ht (x) = ρ . − ⊕ V 0 is compact. Now deﬁne a Note that kt := (P − + P 0 )L−1 t Kt : W → V map Tt : W → W by letting

Tt (x) = Tt (x− + x0 + se) = x− + x0 + kt (x) + ht (x)e = x + kt (x) + ht (x) − s e .

8. Linking

135

Observe that Tt is of the form id + compact. Moreover, the condition ht (∂Q) ∩ Sρ+ = ∅ translates into the condition Tt (x) = ρe for x ∈ ∂Q. Hence the LeraySchauder degree (see for instance Deimling [1; 2.8.3, 2.8.4]) of Tt on Q with respect to ρe is well-deﬁned and, in fact, deg(Tt , Q, ρe) = deg(T0 , Q, ρe) = 1 , as desired.

Let ∇E denote the gradient of E with respect to the scalar product in V . Since ∇E has linear growth, the gradient ﬂow Φ: V × [0, ∞[→ V given by ∂ Φ(x, t) = −∇E Φ(x, t) ∂t Φ(x, 0) = x exists globally. Note that E is non-increasing along ﬂow-lines with % % % %2 d % . E Φ(x, t) = −% ∇E Φ(x, t) % % dt

(8.10)

8.11 Lemma. Φ(·, T ) ∈ Γ for any T ≥ 0. Proof. Clearly ht = Φ(·, t) for 0 ≤ t ≤ T is a homotopy of Φ(·, T ) to the identity; moreover by (8.10) E Φ(x, t) ≤ E(x) ≤ 0

for x ∈ ∂Q .

By Lemmata 8.8 and 8.9 therefore ht (∂Q) ∩ Sρ+ = ∅ for all t. Finally, observe that ∇E(x) = −x− + x+ + ∇G(x) ,

(8.11)

where ∇G is compact. Thus the desired form of Φ(x, t) may be read oﬀ the variation of constant formula t −

0

−t +

t

Φ(x, t) = e x + x + e x +

et−s P − + P 0 + e−(t−s) P + ∇G Φ(x, s) ds

0

+ + − 0 0 =: L− t x + Lt x + Lt x + Kt (x) .

It remains to verify the Palais-Smale condition for E.

136

Chapter II. Minimax Methods

8.12 Lemma. E satisﬁes (P.-S.) on V . Proof. Let (xm ) be a (P.-S.)sequence for E. By (8.11) and Proposition 2.2 it suﬃces to show that (xm ) is bounded in V . Suppose by contradiction that and note that xm → ∞. Normalize zm = xxm m

(8.12)

(−P − + P + )zm −

∇G(xm ) →0. xm

m) We claim that the sequence ym := ∇G(x

xm is precompact. Indeed, by (8.12), the sequence (ym ) is bounded and we may assume that ym → y weakly in H 1/2,2 and strongly in L2 . Moreover, by (8.6), for φ ∈ H 1/2,2 (S 1 ) we may estimate % % 1 ˜ m ), φ % xm % ∇G(xm ) ∇ H(x % % = ≤ c 1 + , φ dt % xm % 2 φL2 . xm xm 0 V L

Choosing φ = φm = ym − y above, it then follows that ym → y strongly in H 1/2,2 as m → ∞. Hence, from (8.12) it follows that zm → z in H 1/2,2 and ˜ grows linearly, passing to the limit in almost everywhere in [0, 1]. Since ∇H the expression 8 7 3 1 2 ˜ xm zm , φ ∇H φ , DE(xm ) = dt −J z˙m , φ − xm xm 0 is allowed by Vitali’s convergence theorem. From (8.6) we thus infer that z satisﬁes the equation z˙ = 3πJ z . d . Hence z ≡ 0. However, 3π does not belong to the spectrum of L = −J dt But since zm → z, and zm = 1, this is impossible. Therefore the original sequence (xm ) must be bounded.

In order to conclude the proof of Theorem 8.6 we can now repeat the argument of Theorem 8.4. Deﬁne β = inf sup E Φ(x, t) . t≥0 x∈Q

From Lemmata 8.10 and 8.11 it follows that Φ(Q, t) ∩ Sρ+ = ∅ for all t ≥ 0. Hence β ≥ α > 0. Suppose by contradiction that β is a regular value for E. By (P.-S.) there exists δ > 0 such that ∇E(x) ≥ δ for all x such that |E(x) − β| < δ. Choose t0 > 0 such that sup E Φ(x, t0 ) < β + δ . x∈Q

9. Parameter Dependence

137

Then by deﬁnition of β, choice of δ, and (8.10), for t ≥ t0 the number sup E Φ(x, t) x∈Q

% % % % % is achieved only at points where %∇E Φ(x, t) % % ≥ δ. Hence by (8.10) d dt

sup E Φ(x, t) ≤ −δ 2 , x∈Q

which gives a contradiction to the deﬁnition of β after time T ≥ t0 + 1δ . The proof is complete. In the next section we will see that by a slight reﬁnement of the above linking argument one can obtain periodic solutions of (8.5) not only with energy arbitrarily close to a given value β0 , but, in fact, for almost every energy level in a suitable neighborhood of β0 – a vast abundance of periodic solutions!

9. Parameter Dependence The existence result for periodic solutions of Hamiltonian systems given in the previous section can be considerably improved, if combined with the ideas presented in Section I.7. In fact, we have the following result of Struwe [21]. 9.1 Theorem. Let H ∈ C 2 (IR2n , IR) and suppose that 1 is a regular value of H and the energy level surface S = S1 = {x; H(x) = 1} is compact and connected. Then there is a number δ0 > 0 such that for almost every β ∈] 1 − δ0 , 1 + δ0 [ there exists a periodic solution x of (8.5) with H(x(t)) ≡ β. In fact, δ0 > 0 may be any number as determined in the remark following Theorem 8.6 such that for β ∈] 1 − δ0 , 1 + δ0 [ the level surface Sβ = H −1 (β) is C 2 -diﬀeomorphic to S. For the proof of Theorem 9.1 we may modify the Hamiltonian H as in the preceding section while preserving the level surfaces Sβ , |1 − β| < δ0 . However, ˜ as follows. Fix any number β0 ∈] 1 − δ0 , 1 + δ0 [ and let now we shift H, 0 < δ < δ0 − |1 − β0 | /3. If suﬃces to show that (8.5) has a periodic solution with energy β for almost every β ∈] β0 − δ , β0 + δ [=: I0 . Let U = H −1 [1 − δ0 + δ, 1 + δ0 − δ] ∼ = S × [−1, 1], IR2n \ U = A ∪ B, 3 2 2 γ = diam U , γ < r < 2γ, 2 πr < b < 2πr as in Section 8 above and choose smooth functions f and g as in that section such that f (s) = 0 for s ≤ −δ, f (s) = b for s ≥ δ, f (s) > 0 for |s| < δ, and where g satisﬁes (8.6). For β ∈ I0 , m ∈ IN, then deﬁne

138

Chapter II. Minimax Methods

⎧0 , ⎪ ⎨ f m(H(x) − β) , ˜ Hβ,m (x) = ⎪ b , ⎩ g |x| ,

if if if if

x∈B , x∈U , x ∈ A, |x| ≤ r , |x| > r ,

satisfying (8.7), and for x ∈ V = H 1/2,2 (S 1 ; IR2n ) let 1 1 1 ˜ β,m (x) dt Eβ,m (x) = H x, ˙ Jx dt − 2 0 0 = A(x) − Gβ,m (x) . Then, as in Section 8, the functional Eβ,m is of class C 2 and satisﬁes (P.-S.)on V . Moreover, critical points x ∈ V of Eβ,m with Eβ,m (x) > 0 correspond to periodic solutions y of (8.5) with energy |H(y) − β| < δ/m; compare Lemma 8.7. More precisely, we have 9.2 Lemma. Let x ∈ V be a critical point of Eβ,m with Eβ,m (x) > 0. Then H x(t) ≡ h ∈ IR with |h − β| < δ/m, and letting >0 , T (x) = mf m H(x) − β the function y(t) = x t/T (x) is a T (x)-periodic solution of (8.5). Proof. The ﬁrst assertion follows just as in the proof of Lemma 8.9. The second assertion is immediate from the equation ˜ β,m (x) = mf m H(x) − β J ∇H(x) x˙ = J ∇H (9.1) = T (x)J ∇H(x) .

Moreover, we have (compare Lemma 8.8): 9.3 Lemma. α > 0, ρ > 0 such that Eβ,m (x) ≥ α for There exist numbers x ∈ Sρ+ = x ∈ V + ; x = ρ , uniformly in m ∈ IN, β ∈ I0 . Proof. Let r0 > 0 be such that Br0 (0; IR2n ) ⊂ B. By Tchebychev’s inequality, we have 1 ≤ r0−2 L1 t; |x(t)| > r0 |x(t)|2 dt ≤ Cx2 0

for x ∈ V . There exists a constant C0 such that ˜ β,m (x) ≤ C0 |x|2 H

for all x ,

9. Parameter Dependence

139

uniformly in m ∈ IN, β ∈ I0 . Since V → L4 [0, 1]; IR2n , moreover, from H¨ olders inequality we obtain 1 2 ˜ Hβ,m x(t) dt ≤ C0 Gβ,m (x) = |x(t)| dt {t;|x(t)|≥r0

0

1/2 ≤ C0 L1 t; |x(t)| ≥ r0

1

|x(t)|4 dt

1/2

≤ Cx3 .

0

On the other hand, for x ∈ V + there holds A(x) =

1 x2 , 2

and the claim follows. In addition, there exists a uniform number R > ρ such that Eβ,m |∂Q ≤ 0 for all β ∈ I0 , m ∈ IN, where Q is deﬁned as in Section 8; compare Lemma 8.9. Finally, the gradient ﬂow Φβ,m (·, t) for Eβ,m for any t ≥ 0 is a member of Γ , for each β ∈ I0 , m ∈ IN. Hence, as in Section 8, for each β ∈ I0 , m ∈ IN there exists a periodic solution xβ,m of (9.1) with Eβ,m (xβ,m ) = inf sup Eβ,m h(x) =: νm (β). h∈Γ x∈Q

We now use variations of β to obtain a-priori bounds on T (xβ,m ) that allow passing to the limit m → ∞ in (9.1) for a suitable sub-sequence and suitable values of β. Observe that, since f ≥ 0, for any ﬁxed x ∈ V , m ∈ IN the map β → Eβ,m (x) is non-decreasing in β ∈ I0 ; in fact, for any x ∈ V we have 1 ∂ dt . Eβ,m (x) = m f m H(x(t)) − β ∂β 0 In particular, if x ∈ V is critical for Eβ,m with Eβ,m (x) > 0, by Lemma 9.2 this yields ∂ Eβ,m (x) = T (x) . ∂β As a consequence of the monotonicity of Eβ,m , clearly also the map β → νm (β) is non-decreasing for any m ∈ IN. Hence this map is almost everywhere ∂ νm ≥ 0, and there holds diﬀerentiable for any m with ∂β

δ

−δ

∂ νm dβ ≤ sup νm (β) − α ≤ sup sup Eβ,m (x) − α = C < ∞ , ∂β m,β m,β x∈Q

140

Chapter II. Minimax Methods

uniformly in m ∈ IN, with α as determined in Lemma 9.3. But then also ∂ νm ∈ L1 [−δ, δ] , lim inf m→∞ ∂β and by Fatou’s lemma we obtain δ δ ∂ ∂ νm dβ ≤ lim inf νm ≤ C . lim inf m→∞ m→∞ ∂β ∂β −δ −δ ∂ In particular, lim inf m→∞ ∂β νm (β) < ∞ for almost every β ∈ I0 . Fix such a β and let Λ ⊂ IN be a subsequence such that

∂ ∂ νm (β) → lim inf νm (β) = Cβ < ∞ m→∞ ∂β ∂β as m → ∞, m ∈ Λ. We may assume

∂ ∂β νm (β)

≤ Cβ + 1 for all m ∈ Λ.

We claim: 9.4 Lemma. For any m ∈ Λ there exists a critical point xm of Eβ,m such that ∂ Eβ,m (xm ) ≤ Cβ + 4. Eβ,m (xm ) = νm (β) and T (xm ) = ∂β Proof: Choose a sequence βk β. We claim there exists a (P.-S.)-sequence (xkm ) for Eβ,m such that ∇Eβ,m (xkm ) → 0 as k → ∞ and such that (9.2)

νm (β) − 2(βk − β) ≤ Eβ,m (xkm ) ≤ Eβk ,m (xkm ) ≤ νm (βk ) + (βk − β) ≤ νm (β) + (Cβ + 2)(βk − β)

for large k. This will imply the assertion of the lemma: By (P.-S.), a subsequence (xkm ) as k → ∞ will accumulate at a critical point xm of Eβ,m , satisfying Eβ,m (xm ) = νm (β) and Eβk ,m (xkm ) − Eβ,m (xkm ) k→∞ βk − β 1 βk mf m(H(xkm ) − β ) dβ dt = lim inf k→∞ βk − β 0 β 1 mf m(H(xm ) − β) dt = T (xm ) . =

Cβ + 4 ≥ lim inf

0

Negating the above claim, suppose there is no (P.-S.)-sequence for Eβ,m satisfying (9.2). Then there exists ε > 0 and k0 ∈ IN such that for all x ∈ V satisfying (9.2) for some k ≥ k0 there holds (9.3)

∇Eβ,m (x)2 ≥ 2ε .

We may assume k0 = 1. Let 0 ≤ ϕ ≤ 1 be a Lipschitz continuous function such that ϕ(s) = 0 for s ≤ 0, ϕ(s) = 1 for s ≥ 1 and deﬁne a family of vector ﬁelds ek by letting

9. Parameter Dependence

141

ek (x) = − 1 − ϕk (x) ∇Eβk ,m (x) + ϕk (x)∇Eβ,m (x) = x− − x+ + 1 − ϕk (x) ∇Gβk ,m (x) + ϕk (x)∇Gβ,m (x) , where

ϕk (x) = ϕ

Eβ,m (x) − νm (β) − 2(βk − β) . βk − β

Let Φk : V × [ 0, ∞[ → V be the corresponding ﬂows, satisfying ∂ Φk (x, t) = ek Φk (x, t) , ∂t Φk (x, 0) = x for all x ∈ V, t ≥ 0. We claim that Φk (·, t) ∈ Γ for any t ≥ 0. Indeed, we may assume νm (β) ≥ 2(βk − β) for all k. Thus, if Eβk ,m (x) ≤ 0 for some x ∈ V , it follows that ϕk (x) = 0, ek (x) = −∇Eβk ,m (x), and hence Eβk ,m Φk (x, t) ≤ 0 for all t ≥ 0. Since this applies, in particular, to any point x ∈ ∂Q, it follows that Φk (∂Q, t) ∩ Sρ+ = ∅ for all t ≥ 0. The remaining properties deﬁning the class Γ are veriﬁed as in the proof of Lemma 8.11. Choose h ∈ Γ such that sup Eβk ,m h(x) ≤ νm (βk ) + (βk − β)

(9.4)

x∈Q

and consider a point x ∈ h(Q). If Eβ,m (x) < νm (β) − 2(βk − β) , by deﬁnition of ek we have ek (x) = −∇Eβk ,m (x) and hence

ek (x), ∇Eβk ,m (x)

V

≤0

for such x. If, on the other hand, Eβ,m (x) ≥ νm (β) − 2(βk − β) , by assumption (9.4) estimate (9.2) is veriﬁed for x. Observe that

ek (x), ∇Eβk ,m (x)

V

≤ −ϕk (x) ∇Eβ,m (x), ∇Eβk ,m (x) V .

But, by (9.3), for points x satisfying (9.2) we can estimate

142

Chapter II. Minimax Methods

∇Eβ,m (x),∇Eβk ,m (x)

V = ∇Eβ,m (x)2 − ∇Eβ,m (x), ∇Gβk ,m (x) − ∇Gβ,m (x) V 1 1 ≥ ∇Eβ,m (x)2 − ∇Gβk ,m (x) − ∇Gβ,m (x)2 2 2

≥ε−C

{t;|x(t)|≤r}

˜ β ,m (x) − ∇H ˜ β,m (x)|2 dt |∇H k

≥ ε − o(1) , where o(1) → 0 as k → ∞ by dominated convergence, uniformly in x. Hence, for large k, for any x ∈ h(Q) we have d Eβ ,m Φk (x, t) |t=0 = ek (x), ∇Eβk ,m (x) V ≤ 0 . dt k In particular, letting ht = Φk (·, t) ◦ h for t ≥ 0, estimate (9.4) holds for any ht , t ≥ 0. Moreover, ht ∈ Γ for any t ≥ 0. Hence, by deﬁnition of νm (β), for any t ≥ 0 we have (9.5) M (t) = sup Eβ,m ht (x) ≥ νm (β) . x∈Q

Together with (9.4), this implies that M (t) is achieved only at points x ∈ ht (Q) satisfying (9.2). Moreover, ϕk (x) = 1 at such points and hence ek (x), ∇Eβ,m(x) V = −Eβ,m (x)2 ≤ −2ε by (9.3). Thus, M (t) is strictly decreasing with d M (t) ≤ −2ε < 0 , dt contradicting (9.5) for suﬃciently large t > 0. This completes the proof. Proof of Theorem 9.1: For β ∈ I0 with lim inf m→∞

∂ νm (β) = Cβ < ∞ ∂β

let Λ and (xm )m∈Λ be as in Lemma 9.4. For any m the function xm solves (9.1) with T (xm ) ≤ Cβ + 4 and satisﬁes |H(xm (t)) − β| <

δ . m

Hence the sequences (xm ), (x˙ m ) are equi-bounded and equi-continuous; By the theorem of Arz´ela-Ascoli, therefore, a subsequence converges C 1 uniformly to a 1-periodic solution x of x˙ = T J ∇H(x)

10. Critical Points of Mountain Pass Type

143

for some T ≥ 0, with H x(t) ≡ β. Moreover, A(xm ) ≥ Eβ,m (xm ) ≥ α; hence A(x) ≥ α > 0. In particular, x is non-constant, T > 0, and y(t) = x(t/T ) is a T -periodic solution of (8.5) on Sβ .

9.5 Notes. (1◦ ) Recently, Ginzburg [1] and Herman [1] have given an example of a smooth Hamiltonian H, possessing a smooth, compact energy level surface carrying no periodic orbit. Theorem 9.1 therefore is best possible. A related result is due to Kuperberg [1]. (2◦) Further applications of the above method are given in Struwe [16], [17], Ambrosetti-Struwe [2]. However, an abstract statement would be quite cumbersome and in each instance, features that are particular to the given problem come into play. Common to all the above applications is a family of functionals Eβ ∈ C 1 (V ) depending monotonically on β ∈ IR. A key technical point is that a pseudo-gradient ﬂow for Eβ should be a pseudo-gradient ﬂow also for Eβk near points x ∈ V such that |Eβk (x) − Eβ (x)| ≤ C|βk − β|, when |βk − β| $ 1.

10. Critical Points of Mountain Pass Type Diﬀerent critical points of functionals E sometimes can be distinguished by the topological type of their neighborhoods in the sub-level sets of E. This is the original idea of M. Morse which led to the development of what is now called Morse theory; see the Introduction. Hofer [2] has observed that such information about the topological type in some cases is available already from the minimax characterization of the corresponding critical value. 10.1 Definition. Let E ∈ C 1 (V ), and suppose u is a critical point of E with E(u) = β. We say u is of mountain pass type if for any neighborhood N of u the set N ∩ Eβ is disconnected. With this notion available we can strengthen the assertion of Theorem 6.1 as follows. For convenience we recall 6.1 Theorem. Suppose E ∈ C 1 (V ) satisfies (P.-S.). Suppose (1◦ ) E(0) = 0 . (2◦ ) ∃ρ > 0, α > 0 : u = ρ ⇒ E(u) ≥ α. (3◦ ) ∃u1 ∈ V : u1 ≥ ρ and E(u1 ) < α. Define Γ = p ∈ C 0 ([0, 1]; V ) ; p(0) = 0, p(1) = u1 . Then

β = inf sup E(u) p∈Γ u∈p

is a critical value.

Now we assert:

144

Chapter II. Minimax Methods

10.2 Theorem. Under the hypotheses of Theorem 6.1 the following holds: (1◦) either E admits a relative minimizer u = 0 with E(u) = β, or (2◦) E admits a critical point u of mountain pass type with E(u) = β.

Remark. Simple examples on IR show that in general also case (1◦ ) occurs; see Figure 10.1.

Fig. 10.1. A function possessing no critical point of mountain pass type

Another variant of Theorem 6.1 is related to results of Chang [4] and Pucci-Serrin [1], [2]. 10.3 Theorem. Suppose E ∈ C 1 (V ) satisﬁes (P.-S.). Suppose 0 is a relative minimizer of E with E(0) = 0, and suppose that E admits a second relative minimizer u1 = 0. Let Γ and β be deﬁned as in Theorem 6.1. Then, (1◦) either there exists a critical point u ∈ Kβ which is not of minimum type, or (2◦) the origin and u1 can be connected in any neighborhood of the set of relative minimizers u of E with E(u) = 0. Necessarily then, β = E(u1 ) = E(0) = 0. The proofs of these results are quite similar in spirit. Proof of Theorem 10.3. Let Γ and β be deﬁned as in Theorem 6.1. Suppose that Kβ consists entirely of relative minimizers of E. Then for each u ∈ Kβ there exists a neighborhood N (u) such that E(u) = β ≤ E(v)

for all v ∈ N (u) .

˜ of Kβ let ε > 0 and Φ be N (u), and for any neighborhood N ˜ . Choosing determined according to Theorem 3.4 for ε = 1 and N = N0 ∩ N p ∈ Γ such that p([0, 1]) ⊂ Eβ+ε , by (1◦ ),(3◦ ) of Theorem 3.4 the path p = Φ(·, 1) ◦ p ∈ Γ satisﬁes Let N0 =

u∈Kβ

p ([0, 1]) ⊂ Φ(Eβ+ε , 1) ⊂ N ∪ Eβ−ε ⊂ N0 ∪ Eβ−ε .

10. Critical Points of Mountain Pass Type

145

But N0 and Eβ−ε by construction are disjoint, hence disconnected. Thus either p ([0, 1]) ⊂ N or p ([0, 1]) ⊂ Eβ−ε . Since the latter contradicts the deﬁnition of ˜ , whence p(0) = 0 and p(1) = u1 can be β we conclude that p ([0, 1]) ⊂ N ⊂ N ˜ of Kβ , as claimed. In particular, 0 ∈ Kβ , connected in any neighborhood N u1 ∈ Kβ , whence β = E(0) = 0 and also E(u1 ) = 0. Proof of Theorem 10.2. Suppose, by contradiction, that Kβ contains no relative minimizers nor critical points of mountain pass type. Then any u ∈ Kβ possesses a neighborhood N (u) such that N (u) ∩ Eβ is non-empty and (path-) connected. Moreover, Kβ ⊂ Eβ . Now we have the following topological lemma; see Hofer [3; Lemma 1]: 10.4 Lemma. Let (M, d) be a metric space and let K and Λ be non-empty subsets of M such that K is compact, Λ is open, and K ⊂ Λ, the closure of Λ. Suppose {N (u) ; u ∈ K} is an open cover of K such that u ∈ N (u) and N (u) ∩ Λ is connected for each u ∈ K. Then there exists a ﬁnite, disjoint open cover {U1 , . . . , UL } of K such that Ul ∩ Λ for each l is contained in a connected component of Λ. Proof. Choose δ > 0 such that for any u ∈ K we have Bδ (u) ⊂ N (u)

(10.1)

for some u ∈ K .

For instance, we may choose a ﬁnite subcover {N (ui ) ; 1 ≤ i ≤ I} of {N (u) ; u ∈ K} and let δ = min max dist(u, M \ N (ui )) > 0 . u∈K 1≤i≤I ∗

Deﬁne an equivalence relation ∼ on K by letting There exist a number m ∈ IN and points ui ∈ K, 0 ≤ i ≤ ∗ u ∼ u ⇔ m + 1, such that u0 = u, um+1 = u and d(ui , ui+1 ) < δ for 0 ≤ i ≤ m. Since K is compact there are only ﬁnitely many equivalence classes, say K1 , . . . , KL . Let δ , Ul = x ∈ M ; dist(x, Kl ) < 4

1≤l≤L.

L Then it is immediate that Uk ∩ Ul = ∅ if k = l and l=1 Ul ⊃ K. It remains to show that each Ul∗ = Ul ∩ Λ is contained in a connected component of Λ. Deﬁne another equivalence relation ∼ on Λ by letting v∼w

⇔

v and w belong to the same connected component of Λ.

146

Chapter II. Minimax Methods

Fix l and let v, w ∈ Ul ∩ Λ. We wish to show that v ∼ w. By deﬁnition of Ul and Kl there exists a ﬁnite chain ui ∈ Kl , 0 ≤ i ≤ m + 1, such that d(v, u0 ) <

δ , 4

d(w, um+1 ) <

δ , 4

and d(ui , ui+1 ) < δ , for 0 ≤ i ≤ m . Set ε = δ − max1≤i≤m d(ui , ui+1 ) > 0. Since K ⊂ Λ, for each 0 ≤ i ≤ m + 1 there exists vi ∈ Λ such that d(ui , vi ) < 2ε < δ, whence also d(vi , vi+1 ) < δ

for 0 ≤ i ≤ m .

But now by (10.1) we have v, v0 ∈ Bδ (u0 ) ∩ Λ ⊂ N (u0 ) ∩ Λ for some u0 ∈ K, and our assumption about the cover {N (u) ; u ∈ K} implies that v ∼ v0 . Similarly, w ∼ vm+1 . Finally, for 0 ≤ i ≤ m we have d(ui , vi+1 ) ≤ d(ui , ui+1 ) + d(ui+1 , vi+1 ) < δ , whence

vi , vi+1 ∈ Bδ (ui ) ∩ Λ ⊂ N (ui ) ∩ Λ ,

and vi ∼ vi+1 for all i = 0, . . . , m. In conclusion v ∼ v0 ∼ v1 ∼ . . . ∼ vm+1 ∼ w , and the proof is complete Proof of Theorem 10.2 (completed). Let {U1 , . . . , UL } be a disjoint open cover L of Kβ as in Lemma 10.4 and set N = l=1 Ul . Choose ε = α > 0 and let ε > 0, Φ be determined according to Theorem 3.4, corresponding to β, ε, and N . Let p ∈ Γ satisfy p([0, 1]) ⊂ Eβ+ε . Then p = Φ(·, 1) ◦ p ∈ Γ and p ([0, 1]) ⊂ Φ(Eβ+ε , 1) ⊂ Eβ−ε ∪ N = Eβ−ε ∪ U1 ∪ . . . ∪ UL . Smoothing p if necessary, we may assume p ∈ C 1 . Choose γ ∈]β −ε, β[ such that γ is a regular value of E ◦ p . Let 0 < t1 < t2 < . . . < t2k−1 < t2k < 1 denote the successive pre-images L of γ under E ◦p , and let Ij = [t2j−1 , t2j ], 1 ≤ j ≤ k. Note that p |Ij ⊂ l=1 Ul . Since the latter is a union of disjoint sets, for any j there is l ∈ {1, . . . , L} such that p (Ij ) ⊂ Ul . But E ◦ p (∂Ij ) = γ < β and Ul ∩ Eβ is connected. Hence we may replace p |Ij by a path p˜: Ij → Ul ∩ Eβ with endpoints p˜|∂Ij = p|∂Ij , for any j = 1, . . . k. In this way we obtain a path p˜ ∈ Γ such that supu∈p˜ E(u) < β, which yields the desired contradiction.

10. Critical Points of Mountain Pass Type

147

Multiple Solutions of Coercive Elliptic Problems We apply these results to a semilinear elliptic problem. Let Ω be a bounded domain in IRn , g: IR → IR a continuous function. For λ ∈ IR consider the problem (10.2) (10.3)

−Δu = λu − g(u) u=0

in Ω , on ∂Ω ,

Let 0 < λ1 < λ2 ≤ λ3 . . . denote the eigenvalues of −Δ with homogeneous Dirichlet boundary condition. Then we may assert 10.5 Theorem. Suppose g is locally Lipschitz with g(0) = 0 and assume that is non-decreasing with the map u → g(u) |u| (1◦) limu→0 g(u) u = 0, and = ∞. (2◦) lim|u|→∞ g(u) u Then for any λ > λ2 problem (10.2), (10.3) admits at least three distinct nontrivial solutions. Note that by assumption (1◦ ) the problem (10.2), (10.3) for any λ ∈ IR admits u ≡ 0 as (trivial) solution. Moreover, by (2◦ ) , the functional E related to (10.2), (10.3) is coercive. The latter stands in contrast with the examples studied earlier in this chapter. As a consequence, the existence of multiple solutions for problem (10.2), (10.3) heavily depends on the behavior of the functional E near u = 0, governed by the parameter λ, whereas in previous examples the behavior near ∞ had been responsible for the nice multiplicity results obtained. Theorem 10.5 is due to Ambrosetti-Mancini [1] and Struwe [5]. Later, Ambrosetti-Lupo [1] were able to simplify the argument signiﬁcantly, and we shall basically follow their approach in the proof below. See also Rabinowitz [11; Theorem 2.42], Chang [1; Theorem 3], and Hofer [1] for related results. If g is odd, then for λk < λ ≤ λk+1 problem (10.2), (10.3) possesses at least k pairs of distinct non-trivial solutions, see for example Ambrosetti [1]. (This can also be deduced from a variant of Theorem 5.7 above, applied to the sub-level set M = E0 of the functional E related to (10.2), (10.3). Note that M is forwardly invariant under the pseudo-gradient ﬂow for E and hence the trajectories of this ﬂow are complete in forward time. Moreover, for λk < λ ≤ λk+1 by (1◦ ) it follows that the genus γˆ (M ) ≥ k.) Without any symmetry assumption optimal multiplicity results for (10.2), (10.3) are not known. However, results of Dancer [2] suggest that in general even for large λ one can expect no more than four non-trivial solutions. ≥ λ}. Then we may Proof of Theorem 10.5. Let u± = min{u > 0 ; g(±u) ±u replace g by the truncated function

g(u), −u− ≤ u ≤ u+ gˆ(u) = λu, u < −u− or u+ < u.

148

Chapter II. Minimax Methods

Observe that if u satisﬁes

−Δu = λu − gˆ(u) =

λu − g(u), if u ∈ [−u− , u+ ] 0, else

then by the weak maximum principle in fact u satisﬁes (10.2), (10.3). Thus we may assume that |g(x, u) − λu| ≤ c and λu − g(u) u ≥ 0 for all u ∈ IR. Hence the functional 1 |∇u|2 − λ|u|2 dx + E(u) = G(x, u) dx 2 Ω Ω is well-deﬁned, and E ∈ C 1 H01,2 (Ω) . Moreover, (P.-S.) is satisﬁed. In fact, E is coercive, and, if we identify H01,2 (Ω) with its dual via the inner product, DE is of the form id + compact. Therefore, (P.-S.) follows from Proposition 2.2. In a ﬁrst step we now want to exhibit a positive solution u ¯ of (10.2), (10.3), and a negative solution u, respectively. This can be done in various ways by using the methods outlined in the previous chapters. For instance, we might truncate the nonlinearity λu − g(u) below or above 0 and proceed as in the proof of Theorem 6.2. However, we can also use the trivial solution u = 0 as a sub-(super-) solution to problem (10.2), (10.3) and minimize the functional E in the cone of non-negative (non-positive) functions, as we did earlier in Sections I.2.3, I.2.4. We choose this latter approach. Let u, u minimize E in M = {u ∈ H01,2 (Ω) ; u ≥ 0}, respectively M = {u ∈ H01,2 (Ω) ; u ≤ 0}. Then, since 0 is a trivial solution to (10.2), (10.3), by Theorem I.2.4 the functions u, u also solve (10.2), (10.3). In particular, u, u ∈ C 2,α (Ω) for some α > 0; see Appendix B. Moreover, if λ > λ1 and if ϕ1 > 0, ϕ1 L2 = 1, denotes a normalized eigenfunction corresponding to λ1 , we have 1 E(εϕ1 ) = (λ1 − λ)ε2 − o(ε2 ) , 2 where o(s)/s → 0 as s → 0, and this is < 0 for suﬃciently small |ε| = 0; hence u, u ≡ 0, and in fact, by the strong maximum principle (Theorem B.4 of Appendix B), the functions u, u cannot have an interior zero. We claim: u and u are relative minimizers of E in H01,2 (Ω). It suﬃces to show that for suﬃciently small ρ > 0 we have E(u) ≤ E(u) for all u ∈ M ρ := {u ∈ H01,2 (Ω) ; (u − u)− H 1,2 ≤ ρ} , 0

where (s)± = ± max{±s, 0}. Fix ρ > 0 and let uρ ∈ M ρ be a minimizer of E in Mρ . Then v, DE(uρ ) ≥ 0 for all v ≥ 0; that is, uρ is a (weak) supersolution to (10.2), (10.3), satisfying −Δuρ ≥ λuρ − g(uρ ) .

10. Critical Points of Mountain Pass Type

149

Choosing v = −(uρ )− ≥ 0 as testing function, we obtain |∇uρ− |2 − λ|uρ− |2 dx + g(uρ− )uρ− dx ≤ 0 . Ω

Ω

◦

But by (1 ) there is a constant c ≥ 0 such that g(u) u ≥ −c|u|2 for all u; thus |∇(uρ− )|2 − (λ + c)|uρ− |2 dx ≤ 0 . Ω

Let Ωρ− = {x ∈ Ω ; uρ < 0}. Then, since u > 0, we have Ln (Ωρ− ) → 0 as ρ → 0. Now by H¨ older’s and Sobolev’s inequalities we can estimate n−2 n n n2 2n 2 ˜ |u| dx ≤ L (Ω) |u| n−2 dx ˜ ˜ Ω Ω n n2 ˜ ≤ C(n) L (Ω) |∇u|2 dx ˜ Ω

˜ and for any Ω ˜ ⊂ Ω. Thus, for all u ∈ H01,2 (Ω),

|∇u|2 dx 1,2

Ω λρ− := inf ; u ∈ H (Ω ) \ {0} → ∞ as ρ → 0 , ρ− 0 |u|2 dx Ω and for ρ > 0 suﬃciently small such that λρ− > (λ + c) we obtain uρ− ≡ 0; that is, uρ ≥ 0. Hence E(uρ ) ≥ E(u), and u is a relative minimizer of E in M ρ , whence also in H01,2 (Ω). The same conclusion is valid for u. Now let Γ = p ∈ C 0 [0, 1]; H01,2(Ω) ; p(0) = u, p(1) = u and denote β = inf sup E(u) . p∈Γ u∈ρ

If β = 0 we are done, because Theorem 10.3 guarantees either the existence of inﬁnitely many relative minimizers or the existence of a critical point u ∈ Kβ not of minimum type and thus distinct from u, u. Since β = 0, this point u must also be distinct from the trivial solution u = 0. If β = 0, Theorem 10.3 may yield the third critical point u = 0. However; by Theorem 10.2 there exists a critical point u with E(u) = β of minimum or mountain pass type. But for λ > λ2 , by assumption (1◦ ) and since the map u → g(u) |u| is non-decreasing, the set E0 where E is negative is connected; thus, u = 0 is not of one of these types, and the proof is complete. Remark. Since in the case β = 0 we have to show only that Kβ = {0}, instead of appealing to Theorem 10.2 it would suﬃce to show that, if u = 0 were the only critical point besides u and u, then 0 can be avoided by a path joining u with u without increasing energy. For λ > λ2 this is easily shown by a direct construction in the spirit of Theorem 10.2.

150

Chapter II. Minimax Methods

Notes 10.6. (1◦ ) In the context of Hamiltonian systems (8.5), Ekeland-Hofer [1] have applied Theorem 10.2 to obtain the existence of periodic solutions with prescribed minimal period for certain convex Hamiltonian functions H; see also Girardi-Matzeu [1], [2]. Similar applications to semilinear wave equations as considered in Section I.6.6 have recently been given by Salvatore [1]. (2◦) Generalizations of Theorem 10.2 to higher-dimensional minimax methods like Theorem 6.3 have been obtained by Bahri-Lions [1], Lazer-Solimini [1], and Viterbo [2]. See Remark 7.3 for an application of these results. Recently, Ghoussoub [1], [2] has presented a uniﬁed approach to results in the spirit of Theorems 10.2 and 10.3.

11. Non-diﬀerentiable Functionals Sometimes a functional E: V → IR∪{±∞} may fail to be Fr´echet-diﬀerentiable on V but may only be Gˆ ateaux-diﬀerentiable on its domain Dom(E) = {u ∈ V ; E(u) < ∞} in direction of a dense space of “testing functions” T ⊂ V . 11.1 Nonlinear scalar field equations: The zero mass case. As an example we consider the problem (11.1)

in IRn , n ≥ 3 ,

−Δu = g(u)

with the asymptotic boundary condition (11.2)

u(x) → 0

(|x| → ∞) .

The associated energy integral is 1 (11.3) E(u) = |∇u|2 dx − G(u) dx , 2 IRn IRn where, as usual, we denote G(u) =

u

g(v) dv 0

a primitive of g. In the case of “positive mass”, that is lim sup g(u)/u ≤ −m < 0 , u→0

similar to the problem studied in Section I.4.1, problem (11.1), (11.2) can be dealt with as a variational problem in H 1,2 (IRn ), where E is diﬀerentiable; see Berestycki-Lions [1]. In contrast to Section I.4.1, however, now we do not exclude the “0-mass case”

11. Non-diﬀerentiable Functionals

g(u) →0 u

151

(u → 0) .

Then a natural space on which to study E is the space D1,2 (IRn ); that is, the closure of C0∞ (IRn ) in the norm u2D1,2 = |∇u|2 dx . IRn

By Sobolev’s inequality, an equivalent characterization is 2n D1,2 (IRn ) = u ∈ L n−2 (IRn ) ; ∇u ∈ L2 (IRn ) . Note that, unless G satisﬁes the condition 2n

|G(u)| ≤ c|u| n−2 , the functional E may be inﬁnite on a dense set of points in this space and hence cannot be Fr´echet-diﬀerentiable on D1,2 (IRn ). To overcome this diﬃculty, in Struwe [6], [7] a variant of the Palais-Smale condition was introduced and a critical point theory was developed, giving rise to existence results for saddle points for a broad class of functionals where standard methods fail. Below, we give an outline of the abstract scheme of this method; then we apply it to our model problem (11.1), (11.2) above to obtain a simple proof for the following result of Berestycki and Lions [1], [2].

u 11.2 Theorem. Suppose g is continuous with primitive G(u) = 0 g(v) dv and satisﬁes the conditions 9 2n (1◦) −∞ ≤ lim supu→0 g(u)u |u| n−2 ≤ 0 , 9 2n (2◦) −∞ ≤ lim sup|u|→∞ g(u)u |u| n−2 ≤ 0 , and suppose there exists a constant ξ1 such that (3◦) G(ξ1 ) > 0 . Moreover, assume that g is odd, that is, (4◦) g(−u) = −g(u) . Then there exist inﬁnitely many radially symmetric solutions ul ∈ D1,2 (IRn ) of (11.1), (11.2) and E(ul ) → ∞ as l → ∞. Remark. Observe that by the maximum function g˜, given by ⎧ ⎨ g(ζ ∗ ), g˜(u) = g(u), ⎩ g(−ζ ∗ ),

principle we may replace g by the if u > ζ ∗ if |u| ≤ ζ ∗ if u < −ζ ∗

where ζ ∗ = inf{ζ ≥ ζ1 ; g(ζ) ≤ 0} ≤ +∞. Indeed, if u solves (11.1), (11.2) for g˜, by the maximum principle u solves (11.1), (11.2) for g. Hence we may assume that instead of (2◦ ) g satisﬁes the stronger hypothesis

152

Chapter II. Minimax Methods

9 2n (5◦) lim|u|→∞ g(u)u |u| n−2 = 0 together with the assumption (6◦) ∃ζ2 > 0 : G(ζ) > 0 for all ζ, |ζ| > ζ2 . For the proof of Theorem 11.2 we now follow Struwe [6]. 11.3 The abstract scheme. Suppose that E: Dom(E) ⊂ V → IR is a densely deﬁned functional on a Banach space V with norm · . Moreover, assume there is a family (TL )L∈IN of Banach spaces T1 ⊂ . . . ⊂ TL ⊂ TL+1 ⊂ . . . ⊂ V with norms · L such that u ≤ . . . ≤ uL+1 ≤ uL

for u ∈ TL ,

and whose union is dense in V : (11.4)

T :=

dense

TL ⊂ V .

L∈IN

(By default, all topological statements refer to the norm-topology of V .) Also suppose that for any u ∈ Dom(E) the restricted functional E|{u}+TL ∈ C 1 (TL ), for any L ∈ IN, and the partial derivative DL E: Dom(E) u → DL E(u) ∈ TL∗ is continuous in the topology of V for any L ∈ IN. We deﬁne u ∈ Dom(E) ⊂ V to be critical if DL E(u) = 0 for all L ∈ IN, and we denote Kβ = {u ∈ Dom(E) ; E(u) = β, DL E(u) = 0, ∀L ∈ IN} the set of critical points with energy β. Suppose that E satisﬁes the following variant of the Palais-Smale condition: : (P.-S.)

Any sequence (um ) in Dom(E) such that E(um ) → β, while DL E(um ) → 0 in TL∗ (m → ∞), for any L ∈ IN, possesses an accumulation point in Kβ .

Note the following:

11. Non-diﬀerentiable Functionals

153

11.4 Lemma. Suppose V satisﬁes (11.4) and E: Dom(E) ⊂ V → IR satisﬁes : (P.-S.). Then for any β ∈ IR the set Kβ is compact and any neighborhood N of Kβ contains a member of the family Nβ,L = {u ∈ Dom(E) ; |E(u) − β| < 1/L, DL E(u)∗L < 1/L}, L > 0 . Moreover, the system Uβ,ρ = {u ∈ V ; ∃v ∈ Kβ : u − v < ρ} is a fundamental system of neighborhoods of Kβ . : any sequence (um ) in Kβ is relatively compact and accuProof. By (P.-S.) mulates at some point u ∈ Kβ . To prove the second assertion, suppose by contradiction that for some neighborhood N of Kβ and any L ∈ IN there is a point uL ∈ Nβ,L \ N . Consider the sequence (uL ). Since for any L we have DL E(uL )∗L ≤ DL E(uL )∗L ≤ 1/L → 0 : we conclude that (uL ) accumulates at a point as L → ∞, L ≥ L , from (P.-S.) u ∈ Kβ , contrary to assumption. The proof for Uβ,ρ is similar. Denote RegL (E) = {u ∈ Dom(E) ; DL E(u) = 0} the set of regular points of E with respect to variations in TL . Then exactly as in Lemma 3.2, using continuity of the partial derivative DL E, we can construct a locally Lipschitz continuous pseudo-gradient vector ﬁeld vL : RegL (E) → TL for E satisfying the conditions vL (u) < 2 min 1, DL E(u)∗L , vL (u), DL E(u) > min 1, DL E(u)∗L DL E(u)∗L . Now the deformation lemma Theorem 3.4 can be carried over easily. Given β ∈ IR, N ⊃ Kβ we determine L ∈ IN, ρ > 0 such that N ⊃ Uβ,ρ ⊃ Uβ,ρ/2 ⊃ Nβ,L . Choose a locally Lipschitz continuous function η, 0 ≤ η ≤ 1, such that η = 0 1 if DL E(u)∗L ≤ 2L and such that η(u) = 1 if DL E(u)∗L ≥ L1 . Thus, in particular, we have η(u) = 1 for u ∈ Nβ,L satisfying |E(u) − β| ≤ L1 . Let ΦL be the ﬂow corresponding to the vector ﬁeld eL (u) = −η(u)vL (u), deﬁned by solving the initial value problem ∂ ΦL (u, t) = eL ΦL (u, t) ∂t ΦL (u, 0) = u

154

Chapter II. Minimax Methods

for u ∈ Dom(E), t ≥ 0. By local Lipschitz continuity and uniform boundedness of eL , the ﬂow ΦL exists globally on Dom(E) × [0, ∞[, is continuous, and ﬁxes critical points of E. Moreover, E is non-increasing along ﬂow-lines, and we have ΦL (Eβ+ε , 1) ⊂ Eβ−ε ∪ N , respectively ΦL (Eβ+ε \ N, 1) ⊂ Eβ−ε ,

1

where ε = min L , 4Lρ 2 ; see the proof of Theorem 3.4. Note that we do not require E to be continuous on its domain. Thus, we have to be careful with truncating the vector ﬁeld eL outside some energy range. However, with this crude tool already, many of our abstract existence results may be carried over. Proof of Theorem 11.2. Let us now implement the above scheme with E given by (11.3) on 1,2 (IRn ) , V = {u ∈ D1,2 (IRn ) ; u(x) = u(|x|)} =: Drad

with norm · = · D1,2 . (We focus on radially symmetric solutions to remove translation invariance.) Also let 1,2 (IRn ) ; u(x) = 0 for |x| ≥ L , TL = u ∈ Drad with norm · L = · = · D1,2 , L ∈ IN. Note that in this way T = L∈IN TL 1,2 simply consists of all functions u ∈ Drad (IRn ) with compact support. Since variations in TL for any L only involve the evaluation of g, respectively G on a compact domain BL (0), it is clear that E(u + ·) is Fr´echet-diﬀerentiable in TL , for any u ∈ Dom(E), any L ∈ IN. Moreover, the diﬀerential DL E: Dom(E) → TL∗ is continuous in the topology of V for any L ∈ IN. 1,2 (IRn ) is represented Note that by radial symmetry any function u ∈ Drad by a function (indiscriminately denoted by u), which is continuous on IRn \ {0}. Moreover, by H¨ olders’s inequality there holds ∞ d |u(r)|2 dr |u(x)|2 ≤ |x| dr n1 n−2 2n ≤2 (11.6)

∞

r (2−n)n−1 dr

|x|

×

≤ C |x| 1,2 (IRn ). for any u ∈ Drad

(2−n)

u , 2

∞

|x| ∞ |x|

2n

|u(r)| n−2 r n−1 dr

d | u(r)|2 r n−1 dr dr

12

11. Non-diﬀerentiable Functionals

155

Decompose 0}

ug = g+ − g− , G = G+ − G− with g± (u)u = max{±g(u)u, 1,2 (IRn ), and G± (u) = 0 g± (v) dv. We assert that, if um u weakly in Drad then

(11.7) IRn

G+ (u) dx = lim

m→∞

IRn

G+ (um ) dx .

Indeed, by (11.6), for any δ > 0 there exists R > 0 such that |um (x)| ≤ δ for |x| ≥ R. Moreover, by assumption (1◦ ) , for any ε > 0 there is δ > 0 such that 2n

G+ (um ) ≤ ε|um | n−2 , if |u| ≤ δ. Hence for large R we can estimate

IRn \BR (0)

G+ (um ) dx ≤ ε

2n

IRn \BR (0)

|um | n−2 dx

2n

≤ C ε um n−2 ≤ C ε , uniformly in m. On the other hand, by assumption (2◦ ) , for any ε > 0 there is a constant C(ε) such that for all u ∈ IR there holds 2n

G+ (u) ≤ ε |u| n−2 + C(ε) . Hence for Ω ⊂ IRn with suﬃciently small measure Ln (Ω) <

G+ (um ) dx ≤ ε Ω

ε , C(ε)

we have

2n

|u| n−2 dx + C(ε)Ln (Ω) ≤ C ε , Ω

uniformly in m, and (11.7) follows by Vitali’s convergence theorem. Since G− ≥ 0, by Fatou’s lemma of course also

IRn

G− (u) dx ≤ lim inf m→∞

IRn

G− (um ) dx ,

and together with (11.7) we obtain that, if um u weakly, then

(11.8) IRn

G(u) dx ≥ lim sup m→∞

IRn

G(um ) dx .

: we need the following estimate reminiscent Moreover, in order to verify (P.-S.) of the “Pohoˇzaev identity”; see Lemma III.1.4.

156

Chapter II. Minimax Methods

11.5 Lemma. Suppose that (1◦ ) and (2◦ ) of Theorem 11.2 hold. Then any 1,2 (IRn ) of equation (11.1) satisﬁes the estimate weak solution u ∈ Drad 2n 2 |∇u| dx ≥ G(u) dx ≥ −∞ . n − 2 IRn IRn Proof. By (2◦ ) and our local regularity result Lemma B.3 of Appendix B, any weak solution to (11.2) is in Lploc and hence also in H 2,p locally, for any p < ∞. Moreover, by (1◦ ),(2◦ ) we have G+ (u) ∈ L1 (IRn ). Testing Equation (11.1) with the function x · ∇u (the generator of the family uR (x) = u(Rx) of dilations of the function u) we may write the product as |∇u|2 n−2 |∇u|2 − div x − ∇u (x · ∇u) = (x · ∇u) Δu 2 2 (11.9) = −(x · ∇u) g(u) = −x · ∇G(u) = − div xG(u) + nG(u) . Integrating over BR (0) and using the radial symmetry of u, we thus obtain that 2n 2 |∇u| dx = G(u) dx n − 2 BR (0) BR (0) 1 2R |∇u|2 + G(u) do − n − 2 ∂BR (0) 2 2n ≥ G(u) dx n − 2 BR (0) 1 2R |∇u|2 + G+ (u) do . − n − 2 ∂BR (0) 2 Since ∇u ∈ L2 (IRn ), G+ (u) ∈ L1 (IRn ), if we let R → ∞ in a suitable way the boundary integral tends to zero. Moreover, |∇u|2 dx → |∇u|2 dx , BR (0) IRn G+ (u) dx → G+ (u) dx , IRn

BR (0)

while from Beppo Levi’s theorem it follows that G− (u) dx ≥ lim sup G− (u) dx . IRn

R→∞

BR (0)

The proof is complete. Remark that under scaling u → uR (x) = u(Rx) the functional E behaves like 1 E(uR ) = R2−n |∇u|2 dx − R−n G(u) dx ; 2 IRn IRn

11. Non-diﬀerentiable Functionals

that is, d 2−n E(uR )|R=1 = dR 2

IRn

157

|∇u|2 dx + n

IRn

G(u) dx .

Hence we may perform a preliminary normalization of admissible functions by restricting E to the set 2n |∇u|2 dx = G(u) dx} M = {u ∈ Dom(E) ; u = 0, n − 2 IRn IRn of functions which are stationary for E with respect to dilations. Note that for u ∈ M we have 1 |∇u|2 dx ; E(u) = n IRn that is, E|M is continuous and coercive with respect to the norm in D1,2 (IRn ). Moreover, E|M satisﬁes the following compactness condition: 11.6 Lemma. Suppose that for a sequence (um ) in M as m → ∞ we have E(um ) → β while DL E(um ) → 0 ∈ TL∗ for any L. Then (um ) accumulates at : on a critical point u ∈ M of E and E(u) = β. That is, E satisﬁes (P.-S.) : M (while it seems unlikely that one can even show boundedness of a (P.-S.) sequence in general). : sequence for E in M . By coerciveness, (um ) is Proof. Let (um ) be a (P.-S.) 1,2 (IRn ), which implies bounded and we may assume that um u weakly in Drad 1 strong convergence g(um ) → g(u) in L (Ω) for any Ω ⊂⊂ IRn . Thus for any ϕ ∈ C0∞ (IRn ) and suﬃciently large L, as m → ∞ we have ∇um ∇ϕ − g(um )ϕ dx ϕ, DL E(um ) = n IR ∇u∇ϕ − g(u)ϕ dx = ϕ, DL E(u) = 0 ; → IRn

1,2 (IRn ) is a critical point of E and hence weakly solves (11.1). that is, u ∈ Drad By Lemma 11.5 and (11.8) therefore 2n 2n lim sup |∇u|2 dx ≥ G(u) dx ≥ G(um ) dx n − 2 IRn n − 2 m→∞ IRn IRn |∇um |2 dx ≥ |∇u|2 dx . = lim sup m→∞

IRn

IRn

Here we have also used the normalization condition for

um ∈ M . Hence um → u 1,2 (IRn ), and also IRn G(um ) dx → IRn G(u) dx; in particular, strongly in Drad u ∈ M and E(um ) → E(u) = β, as claimed. Now we investigate the set M more closely.

158

Chapter II. Minimax Methods

Denote k: IR+ × Dom(E) → IR the mapping 2Rn+1 d E(uρ )|ρ=R 2 − n dρ 2n 2 2 =R |∇u| dx − G(u) dx . n − 2 IRn IRn

k(R, u) = (11.10)

Then for any L ∈ IN and any u ∈ Dom(E) we have k( · , u + · ) ∈ C 1 (IR+ × older’s inequality the partial derivatives ∂L k(R, u) ∈ TL ). Moreover, by H¨ ∂ ∂2 ∗ TL , ∂R k(R, u), and ∂R2 k(R, u) are continuous and uniformly bounded on bounded sets {(R, u) ∈ IR+ × M + TL ; R + u ≤ C}. Finally, for u ∈ M , at R = 1 we have ∂ k(R, u)|R=1 = 2 |∇u|2 dx = 2n E(u) > 0 . n ∂R IR Thus, by the implicit function theorem, for any β > 0 and any L ∈ IN there exists ρ = ρβ,L > 0 and a continuous map R = Rβ,L on a neighborhood Vβ,L = {u ∈ M ; β/2 ≤ E(u) ≤ 2β} + B2ρ (0; TL ) such that R(v+·) ∈ C 1 (TL ) and k R(v), v ≡ 0 for v ∈ Vβ,L ; that is, vR(v) (x) = v R(v)x ∈ M . Denote πβ,L : Vβ,L → M the map v → πβ,L (v) = vR(v) . Remark that πβ,L is continuous. For the next lemma let Kβ = {u ∈ M ; E(u) = β, DL E(u) = 0 for all L} , Nβ,L = {u ∈ M ; |E(u) − β| < 1/L, DL E(u)∗L < 1/L} . Note that by Lemma 11.6 the assertions of Lemma 11.4 hold for Kβ , Nβ,L as above. We now construct a pseudo-gradient ﬂow for E on M . 11.7 Lemma. For any β > 0, any ε > 0 and any neighborhood N of Kβ there exist ε ∈]0, ε[ and a continuous family Φ: M × [0, 1] → M of odd continuous maps Φ( · , t): M → M such that (1◦) Φ(u, t) = u if DL E(u) = 0 for all L ∈ IN, or t = 0, or if |E(u) − β| ≥ ε, (2◦) E Φ(u, t) is non-increasing in t for any u ∈ M , (3◦) Φ(Eβ+ε \ N, 1) ⊂ Eβ−ε . Proof. Choose integers L < L such that N ⊃ Nβ,L ⊃ Nβ,L and let vL : {u ∈ M ; DL E(u) = 0} → TL be an odd, continuous pseudogradient vector ﬁeld for E, satisfying

11. Non-diﬀerentiable Functionals

159

vL (u)L < 2 min 1, DL E(u)∗L , vL (u), DL E(u) > min 1, DL E(u)∗L DL E(u)∗L , for all u ∈ M such that DL E(u) = 0. Let η, ϕ be continuous cut-oﬀ functions 0 ≤ η, ϕ ≤ 1, η(u) = η(−u), η ≡ 0 on Nβ,L , η ≡ 1 oﬀ Nβ,L , ϕ(s) = 0 for 1 |s − β| ≥ 2ε, ϕ(s) = 1 for |s − β| < ε, where ε ≤ min{ 2ε , 4L } will be determined in the sequel. We truncate vL as usual and let eL (u) = −ϕ E(u) η(u)vL (u) . Note that eL : M → TL is odd and continuous. Let ρβ,L , and πβ,L be deﬁned as above. Provided that ε ≤ β4 , for t ≤ ρβ,L then we may let Φ(u, t) = πβ,L u + teL (u) . Φ is continuous, odd, and satisﬁes (1◦ ). Moreover, for ﬁxed u ∈ M the term E Φ(u, t) is diﬀerentiable in t. Indeed, letting R(t) = R u + teL (u) , uR (x) = u Rx for brevity, we have d d E Φ(u, t) t=t = E (u + teL (u))R(t) |t=t0 0 dt dt d d E u + t0 eL (u) R R=R(t ) R(t)|t=t0 = 0 dR dt 7 8 + eL (u) R(t ) , DL E Φ(u, t0 ) .

Since u + t0 eL (u) R=R(t

0

= Φ(u, t0 ) ∈ M the ﬁrst term vanishes. Moreover, the second up to a factor −ϕ E(u) η(u) equals 8 7 vL (u) R(t ) , DL E Φ(u, t0 ) 0 2−n = R(t0 ) ∇ u + t0 eL (u) ∇vL (u) dx n IR −n g u + t0 eL (u) vL (u) dx − R(t0 ) IRn 8 7 =: vL (u), DL E(u) − γ(t0 ) , 0)

where the last line deﬁnes the “error function” γ. d R(t) uniformly bounded on Note that t → R(t) is diﬀerentiable with dt ◦ bounded sets and that by condition (5 ) of the theorem and Vitali’s convergence theorem also 2n g u + teL (u) − g(u) n+2 dx → 0 as t → 0 , BL (0)

uniformly on bounded sets of functions u ∈ M . Hence also the error

160

Chapter II. Minimax Methods

γ(t) ≤ c |R(t) − 1| + |t| +

IRn

g u + teL (u) − g(u) eL (u) dx

→ 0 as t → 0 . In particular, we can achieve that uniformly in u ∈ Eβ+ε we have γ(t) ≤ for 0 ≤ t ≤ t suﬃciently small. By choice of ϕ and η this implies that ϕ E(u) η(u) d E Φ(u, t) ≤ − for u ∈ M, 0 ≤ t ≤ t . dt 2L2 Hence, with ε ≤ t = 1.

t , 4L2

1 2L2

(2◦ ) and (3◦ ) follow. Rescaling time we may assume

Finally, we can conclude the proof of Theorem 11.2. For l ∈ IN let Σl = {A ⊂ M ; A closed, A = −A, γ(A) ≥ l} , where γ denotes the Krasnoselskii genus introduced in Section 5.1, and deﬁne βl = inf sup E(u) . A∈Σl u∈A

11.8 Lemma. (1◦ ) For any l ∈ IN the class Σl is nonvoid, in particular, M = ∅. (2◦) The numbers βl are critical values of E for any l ∈ IN. Moreover, βl → ∞ as l → ∞. Proof. (1◦ ) Fix l ∈ IN. By condition (6◦ ) on G we can ﬁnd an l-dimensional ∞ (IRn ) and a constant α1 > 0 such that for w ∈ S = {w ∈ subspace W ⊂ C0,rad W ; w = α1 } we have IRn

G(w) dx > 0 .

By (11.10) we can ﬁnd τ > 0 such that (11.11)

k(τ, w) < 0

for all w ∈ S. Scaling x with τ , if necessary, we may assume that τ = 1. Since W ⊂ C0∞ there exists another constant α2 such that wL∞ ≤ α2 for w ∈ S. 1,2 (IRn ) given by Consider the truncation mapping δ: W → Drad ⎧ w(x) > α2 , ⎨ α2 , δ(w)(x) = w(x), |w(x)| ≤ α2 , ⎩ −α2 , w(x) < −α2 . Note that δ is continuous and odd. Since the functions δ(w) are uniformly bounded and have uniform compact support, clearly G δ(w) dx ≤ c (11.12) IRn

11. Non-diﬀerentiable Functionals

161

uniformly in w ∈ W . On the other hand, for any w ∈ S there holds |∇δ(μw)|2 dx → ∞ (μ → ∞) , (11.13) IRn

as is easily veriﬁed. For w ∈ W , with w ≥ α1 , let

J(w) = k 1, δ(w)

and extend J continuously to an even map J: W → IR such that J(w) < 0 for w < α1 . (Note that by (11.11) we have J(w) = k(1, w) < 0 for w ∈ S; that is, for w = α1 .) By (11.12), (11.13) the set Ω = {w ∈ W ; J(w) < 0} then is an open, bounded, symmetric neighborhood of 0 ∈ W . Hence, from Proposition 5.2 we deduce that the boundary A of Ω relative to W has genus γ(A) ≥ l. Since the mapping δ is odd and continuous, by supervariance of the genus, Proposition 5.4(4◦ ), also γ δ(A) ≥ l. Moreover, since J(A) = {0} and δ(A) 0, we clearly have δ(A) ⊂ M , concluding the proof of (1◦ ). (2◦) By part (1◦ ) and Lemma 11.7 the numbers βl are well-deﬁned and critical; see the proof of Theorem 4.2. To show that βl → ∞ (l → ∞) assume by contradiction that βl ≤ β uniformly in l. Thus, we can ﬁnd a sequence of sets Al ∈ Σl such that E(u) ≤ 2β for u ∈ Al , l ∈ IN. Letting A = l Al by Proposition 5.4(2◦ ) and Proposition 5.3 there exists an inﬁnite sequence of mutually orthogonal vectors um ∈ A. By coerciveness and uniform boundedness of E on A, there holds um ≤ C for all m, and hence we may extract a weakly convergent subsequence (um ) (relabeled). By mutual orthogonality, um 0 weakly (m → ∞). Decomposing g = g+ − g− as above, with u · g± (u) = max{0, ±u · g(u)}, by (11.7) therefore G+ (um ) dx → 0 (m → ∞) . (11.14) IRn

But for any u ∈ M

2n |∇u| dx ≤ |∇u| dx + n n n −2 IR IR 2n G+ (u) dx . = n − 2 IRn 2

(11.15)

2

IRn

G− (u) dx

1,2 (IRn ) as m → ∞. On the other hand, by Thus, um → 0 strongly in Drad ◦ ◦ assumptions (1 ), (2 ) there exists a constant c > 0 such that for all u there holds 2n g(u)u ≤ c|u| n−2 ,

and consequently 2n

G(u) ≤ c|u| n−2 . Hence, for u ∈ M , by the Sobolev embedding theorem there holds

162

Chapter II. Minimax Methods

u2 =

IRn

|∇u|2 dx =

2n n−2

IRn

2n

2n

G(u) dx ≤ cu n−2 ≤ cu n−2 . 2n L n−2

Dividing by u2 , we conclude that u ≥ c > 0 is uniformly bounded away from 0 for u ∈ M , and a contradiction to (11.14), (11.15) results. This concludes the proof. 11.9 Notes. In a recent paper, Duc [1] has developed a variational approach to singular elliptic boundary value problems which is similar to the method outlined above. Duc only requires continuity of directional derivatives of E. However, in exchange, only a weaker form of the deformation lemma can be established; see Duc [1; Lemma 2.5]. Very interesting new ideas in this regard can also be found in Duc [2]. Lack of diﬀerentiability is encountered in a diﬀerent way when dealing with functionals involving a combination of a diﬀerentiable and a convex term, as in the case of free boundary problems. For such functionals E, using the notion of sub-diﬀerential introduced in Section I.6, a diﬀerential DE may be deﬁned as a set-valued map. Suitable extensions of minimax techniques to such problems have been obtained by Chang [2] and Szulkin [1]. More generally, using the concept of generalized gradients introduced by Clarke [1], [5], Chang [2] develops a complete variational theory also for Lipschitz maps satisfying a (P.-S.)-type compactness condition. In Ambrosetti-Struwe [2] these results, in combination with the technique of parameter variations described in Section II.9 above, are used to establish the existence of steady vortex rings in an ideal ﬂuid for a prescribed, positive, non-decreasing vorticity function. Previously, this problem had been studied by Fraenkel-Berger [1] by a constrained minimization technique; however, their approach gave rise to a Lagrange multiplier that could not be controlled.

12. Lusternik-Schnirelman Theory on Convex Sets In applications we also frequently encounter functionals on closed and convex subsets of Banach spaces. In fact, this is the natural setting for variational inequalities where the class of admissible functions is restricted by inequality constraints; see Sections I.2.3, I.2.4. Functionals on closed convex sets also arise in certain geometric problems, as we have seen in our discussion of the classical Plateau problem in Sections I.2.7–I.2.10. In fact it was precisely for the latter problem, with the aim of re-deriving the mountain pass lemma for minimal surfaces due to Morse-Tompkins [1], [2] and Shiﬀman [2], [3], that variational methods for functionals on closed convex sets were ﬁrst systematically developed; see Struwe [9], [18].

12. Lusternik-Schnirelman Theory on Convex Sets

163

Suppose M is a closed, convex subset of a Banach space V , and suppose that E: M → IR possesses an extension E ∈ C 1 (V ; IR) to V . For u ∈ M deﬁne g(u) =

sup u − v, DE(u)

v∈M u−v<1

as a measure for the slope of E in M . Clearly, if M = V we have g(u) = DE(u). More generally, we obtain 12.1 Lemma. If E ∈ C 1 (V ), the function g is continuous in M . Proof. Suppose um → u (m → ∞), where um , u ∈ M . Then for any v ∈ M such that u − v < 1 and suﬃciently large m there also holds um − v < 1. Hence for any such v ∈ M we may estimate u − v, DE(u) = lim um − v, DE(um ) m→∞

≤ lim sup g(um ) . m→∞

Passing to the supremum with respect to v in this inequality, we infer that g(u) ≤ lim sup g(um ) . m→∞

On the other hand, if for εm 0 we choose vm ∈ M such that um − vm < 1 and um − vm , DE(um ) ≥ g(um ) − εm , by convexity of M , also the vectors wm = um − uum + 1 − um − u vm = um + 1 − um − u (vm − um ) belong to M and satisfy wm − u ≤ um − u + 1 − um − u vm − um < um − u + 1 − um − u = 1 , while vm − wm = um − u vm − um ≤ um − u → 0 . Thus, g(u) ≥ lim supu − wm , DE(u) m→∞

= lim supum − vm , DE(um ) = lim sup g(um ) , m→∞

and the proof is complete.

m→∞

164

Chapter II. Minimax Methods

12.2 Definition. A point u ∈ M is critical if g(u) = 0, otherwise u is regular. If E(u) = β for some critical point u ∈ M of E, the value β is critical; otherwise β is regular. This deﬁnition coincides with the deﬁnition of regular or critical points (values) of a functional given earlier, if M = V . Moreover, as usual we let Mβ = {u ∈ M ; E(u) < β} , Kβ = {u ∈ M ; E(u) = β, g(u) = 0} , Nβ,δ = {u ∈ M ; |E(u) − β| < δ, g(u) < δ} , Uβ,ρ = {u ∈ M ; ∃v ∈ Kβ : u − v < ρ} denote the sub-level sets, critical sets and families of neighborhoods of Kβ , for any β ∈ IR. 12.3 Definition. E satisﬁes the Palais-Smale condition on M if the following is true: Any sequence (um ) in M such that |E(um )| ≤ c uni(P.-S.)M formly, while g(um ) → 0 (m → ∞), is relatively compact. 12.4 Lemma. Suppose E satisﬁes (P.-S.)M . Then for any β ∈ IR the set Kβ is compact. Moreover, the families {Nβ,δ ; δ > 0}, respectively {Uβ,ρ ; ρ > 0} constitute fundamental systems of neighborhoods of Kβ . The proof is identical with that of Lemma 2.3. Denote ˜ = {u ∈ M ; g(u) = 0} M the set of regular points of E, and let ˜ K = {u ∈ M ; g(u) = 0} = M \ M be the set of critical points . ˜ → V is a pseudo-gradient 12.5 Definition. A locally Lipschitz vector ﬁeld v: M vector ﬁeld for E on M if there exists c > 0 such that (1◦) u + v(u) ∈ M, (2◦) v(u) < min{1, g(u)} , (3◦) v(u), DE(u) < −c min 1, g(u) g(u) , ˜. for all u ∈ M Arguing as in the proof of Lemma 3.2 we establish:

166

Chapter II. Minimax Methods

and since M is convex, this initial value problem may be solved by Euler’s method. Assertions (1◦ ), (2◦ ) are trivially satisﬁed by deﬁnition of e, (3◦ ) is proved exactly as in Theorem 3.4.

The results from the preceding sections now may be carried over to functionals on closed, convex sets. In particular, similar to Theorem 10.3 we have: 12.8 Theorem. Suppose M is a closed, convex subset of a Banach space V, E ∈ C 1 (V ) satisﬁes (P.-S.)M on M , and suppose E admits two distinct relative minima u1 , u2 in M . Then either E(u1 ) = E(u2 ) = β and u1 , u2 can be connected in any neighborhood of the set of relative minima u ∈ M of E with E(u) = β, or there exists a critical point u of E in M which is not a relative minimizer of E.

Applications to Semilinear Elliptic Boundary Value Problems Here we will not enter into a detailed discussion of applications of these methods to the Plateau problem for minimal surfaces and for surfaces of constant mean curvature for which they were developed. The reader will ﬁnd this material in the lecture notes of Struwe [18], devoted exclusively to this topic, and in Chang-Eells [1], Jost-Struwe [1], Struwe [13]. Nor will we touch upon applications to variational inequalities. In this context, variational methods ﬁrst seem to have been applied by Miersemann [1] to eigenvalue problems in a cone; see also Kuˇcera [1] and Quittner [1]. Using the methods outlined above, a uniﬁed approach to equations and inequalities can be achieved. Instead, we re-derive Amann’s [2], [3] famous “three solution theorem” on the existence of “unstable” solutions of semilinear elliptic boundary value problems, conﬁned in an order interval between sub- and supersolutions, in the variational case. 12.9 Theorem. Suppose Ω is a bounded domain in IRn and g: IR → IR is of class C 1 , satisfying the growth condition (1◦ )

|gu (u)| ≤ c 1 + |u|p−2 ,

for some p <

Also suppose that the problem (2◦ )

−Δu = g(u)

(3◦ )

u=0

in Ω , on ∂Ω ,

2n . n−2

12. Lusternik-Schnirelman Theory on Convex Sets

167

admits two pairs of sub- and supersolutions u1 ≤ u1 ≤ u2 ≤ u2 ∈ C 2 ∩H01,2 (Ω). Then either u1 or u2 weakly solves (2◦ ), (3◦ ), or (2◦ ), (3◦ ) admits at least three distinct solutions ui , u1 ≤ ui ≤ u2 , i = 1, 2, 3. Proof. By Theorem I.2.4 the functional E ∈ C 1 H01,2 (Ω) related to (2◦ ), (3◦ ), given by 1 |∇u|2 dx − G(u) dx , E(u) = 2 Ω Ω admits critical points ui which are relative minima of E in the order intervals ui ≤ ui ≤ ui , i = 1, 2. Let u = u1 , u = u2 , and deﬁne M = {u ∈ H01,2 (Ω) ; u ≤ u ≤ u a.e.} . Observe that – unless u1 respectively u2 solves (2◦ ),(3◦ ) – u1 and u2 are also relative minima of E in M . To see this, for i = 1, 2 consider Mi = {u ∈ H01,2 (Ω) ; ui ≤ u ≤ ui almost everywhere} and for ρ > 0, as in the proof of Theorem 10.5, let Miρ = {u ∈ M ; ∃v ∈ Mi : u − vH 1,2 ≤ ρ} . 0

Miρ

Note that is closed and convex, hence weakly closed, and E is coercive and weakly lower semi-continuous on Miρ with respect to H01,2 (Ω). By Theorem I.1.2, therefore, E admits relative minimizers uρi ∈ Miρ , i = 1, 2, for any ρ > 0, and there holds 7 ρ 8 7 8 (u1 − u1 )+ , DE(uρ1 ) ≤ 0 ≤ (u2 − uρ2 )+ , DE(uρ2 ) , with (s)+ = max{0, s}, as usual. Subtracting the relations 7 ρ 8 7 8 (u1 − u1 )+ , DE(u1 ) ≥ 0 ≥ (u2 − uρ2 )+ , DE(u2 ) , we obtain that

∇(uρ − u1 )+ 2 dx ≤ ( g(uρ1 ) − g(u1 ) (uρ1 − u1 )+ dx 1 Ω Ω 1 ≤ gu (u1 + s(uρ1 − u1 )) ds (uρ1 − u1 )2+ dx

(uρ1 − u1 )+ 2H 1,2 = 0

0

Ω

≤ sup gu (v) v∈M1ρ

p

L p−2

(uρ1 − u1 )+ 2Lp

≤ C 1 + sup vp−2 Lp · L

n

v∈M1ρ

{x ; uρ1 (x) > u1 (x)}

2γ

· (uρ1 − u1 )+ 2

2n

L n−2

where γ = 1 − n−2 p > 0, and an analogous estimate for (u2 − uρ2 )+ . By 2n Sobolev’s embedding theorem

12. Lusternik-Schnirelman Theory on Convex Sets

165

˜ → V , satisfying 12.6 Lemma. There exists a pseudo-gradient vector ﬁeld v: M (3◦ ) of Deﬁnition 12.5 with c = 12 . ˜ choose w = w(u) ∈ M such that Proof. For u ∈ M (12.1) u − w < min 1, g(u) , 1 u − w, DE(u) > min 1, g(u) g(u) . (12.2) 2 Now, as in the proof of Lemma 3.2, let {Uι ; ι ∈ I} be a locally ﬁnite open ˜ such that for any ι ∈ I and any u ∈ Uι the conditions (12.1), (12.2) cover of M hold with w = w(uι ), for some uι ∈ Uι . Then let {ϕι ; ι ∈ I} be a locally ˜ subordinate to {Uι } and deﬁne Lipschitz partition of unity on M ϕι (u) w(uι ) − u , v(u) = ι∈I

˜ , as claimed. for u ∈ V . The resulting v is a pseudo-gradient vector ﬁeld on M 12.7 Theorem (Deformation Lemma). Suppose M ⊂ V is closed and convex, E ∈ C 1 (V ) satisﬁes (P.-S.)M on M , and let β ∈ IR, ε > 0 be given. Then for any neighborhood N of Kβ there exist ε ∈]0, ε[ and a continuous deformation Φ: M × [0, 1] → M such that (1◦) Φ(u, t) = u if g(u) = 0, or if t = 0, or if |E(u) − β| ≥ ε; (2◦) E Φ(u, t) is non-increasing in t, for any u ∈ M ; (3◦) Φ(Mβ+ε , 1) ⊂ Mβ−ε ∪ N , respectively Φ(Mβ+ε \ N, 1) ⊂ Mβ−ε . Proof. For ε < min{ε/2, δ/4}, where N ⊃ Uβ,ρ ⊃ Uβ,ρ/2 ⊃ Nβ,δ , as in the proof of Theorem 3.4, let η be locally Lipschitz with 0 ≤ η ≤ 1, η(u) = 0 in Nβ,δ/2 , η(u) = 1 on Nβ,δ and choose a smooth cut-oﬀ function 0 ≤ ϕ ≤ 1 such that ϕ(s) = 1 for |s − β| ≤ ε, ϕ(s) = 0 for |s − β| ≥ 2ε. Deﬁne

˜ e(u) = η(u)ϕ E(u) v(u), u ∈ M 0, u∈K . The vector ﬁeld e is Lipschitz continuous, uniformly bounded, and induces a global ﬂow Φ: M × [0, 1] → M such that ∂ Φ(u, t) = e Φ(u, t) ∂t Φ(u, 0) = 0 . Note that since v (and therefore e) satisﬁes the condition ˜ , u + v(u) ∈ M , for all u ∈ M

168

Chapter II. Minimax Methods

vLp ≤ cv

2n

L n−2

≤ c inf

w∈M1

v − w

2n

L n−2

+ w

2n

L n−2

≤ c inf v − wH 1,2 + C ≤ C < ∞ w∈M1

0

2n ≤ c(uρ1 − u1 )+ H 1,2 , whence for v ∈ M1ρ , ρ ≤ 1. Similarly, (uρ1 − u1 )+ n−2 L with a uniform constant C for all ρ > 0 there holds 2γ (12.5) (uρ1 − u1 )+ 2H 1,2 ≤ CLn {x ; uρ1 (x) > u1 (x)} (uρ1 − u1 )+ 2H 1,2 , 0

0

uρ2 )+ .

and an analogous estimate for (u2 − ˜i of E in Mi , i = 1, 2. As ρ → 0 the functions uρi accumulate at minimizers u Arguing as in the proof of Theorem I.2.4, and using the regularity result Lemma B.3 of Appendix B, these functions u ˜i are classical solutions of (2◦ ),(3◦ ). Now, ◦ ◦ ˜2 = if u1 , u2 do not solve problem (2 ),(3 ), then in particular we have u˜1 = u1 , u ˜1 < u1 , u2 < u ˜2 . u2 . From the strong maximum principle we then infer that u Hence Ln {x; uρ1 (x) > u1 (x)} → 0 (ρ → 0) , and from (12.5) it follows that uρ1 ∈ M1 for suﬃciently small ρ > 0, showing that u1 is relatively minimal for E in M . Similarly for u2 . Thus, u1 and u2 are relative minima of E in M . By Theorem 12.8 the functional E either admits inﬁnitely many relative minima in M or possesses at least one critical point u3 ∈ M which is not a relative minimizer of E, hence distinct from u1 , u2 . Finally, observe that any critical point u of E in M weakly solves (2◦ ), ◦ (3 ): Indeed for any ϕ ∈ H01,2 (Ω), ε > 0, let vε = min u, max{u, u + εϕ} = u + εϕ − ϕε + ϕε ∈ M with ϕε = max{0, u + εϕ − u} ≥ 0, ϕε = max{0, u − (u + εϕ)} ≥ 0, as in the proof of Theorem I.2.4. Then, if u ∈ M is critical, as ε → 0 we have u − vε , DE(u) ≤ u − vε H 1,2 0

sup v∈M u−v 1,2 ≤1 H 0

u − v, DE(u)

≤ O(ε)g(u) = 0 . From this inequality the desired conclusion follows exactly as in the proof of Theorem I.2.4.

12.10 Notes. A diﬀerent variational approach to this result was suggested by Chang [5], who used the regularizing properties of the heat ﬂow to reduce the problem to a variational problem on an open subset of a suitable Banach space to which the standard methods could be applied. A combination of topological degree and variational methods was used by Hofer [1] to obtain even higher multiplicity results for certain problems.

Chapter III

Limit Cases of the Palais-Smale Condition

Condition (P.-S.) may seem rather restrictive. Actually, as Hildebrandt [4; p. 324] records, for quite a while many mathematicians felt convinced that inspite of its success in dealing with one-dimensional variational problems like geodesics (see Birkhoﬀ’s Theorem I.4.4, for example, or Palais’ [3] work on closed geodesics), the Palais-Smale condition could never play a role in the solution of “interesting” variational problems in higher dimensions. Recent advances in the Calculus of Variations have changed this view and it has become apparent that the methods of Palais and Smale apply to many problems of physical and/or geometric interest and – in particular – that the Palais-Smale condition will in general hold true for such problems in a broad range of energies. Moreover, the failure of (P.-S.) at certain energy levels reﬂects highly interesting phenomena related to internal symmetries of the systems under study. Geometrically, we might attribute the non-compactness of certain Palais-Smale sequences to the “separation of spheres”, or to “bubbling”, perhaps associated with a “change in topology”. With some twist of imagination, and speaking again in physical terms, we might observe “phase transitions” or “particle creation” at the energy levels where (P.-S.) fails. Such phenomena seem to have ﬁrst been observed by Sacks-Uhlenbeck [1] and – independently – by Wente [5] in the context of two-dimensional harmonic maps, respectively, in the context of surfaces of prescribed constant mean curvature. (See Sections 5 and 6 below.) In these cases the terms “separation of spheres” and “bubbling” have a clear geometric meaning. More recently, Sedlacek [1] has uncovered similar results also for Yang-Mills connections. If interpreted appropriately, very early indications of such phenomena already may be found in the work of Douglas [2], Morse-Tompkins [2] and Shiﬀman [2] on minimal surfaces of higher genus and/or connectivity. In this case, a “change in topology” in fact sometimes may be observed even physically as one tries to realize a multiply connected or higher genus minimal surface in a soap ﬁlm experiment. See Jost-Struwe [1] for a modern approach to these results. Mathematically, it seems that non-compact group actions give rise to these eﬀects. In physics and geometry, of course, such group actions naturally arise as “symmetries” from the requirements of scale or gauge invariance; in particular, in the examples of Sacks-Uhlenbeck and Wente cited above, from conformal invariance.

170

Chapter III. Limit Cases of the Palais-Smale Condition

A symmetry may either be “manifest” or “broken” by interaction terms. As we shall see, existence results in the spirit of Theorem II.6.1 for problems with non-compact internal symmetries often depend on the extent to which the symmetry is broken or perturbed, as measured by whether or not certain mountain-pass energy levels diﬀer from the energy levels where “bubbling” may occur. As in the case of the non-compact minimization problems studied in Section I.4, in many instances the latter question can be answered by comparing the problem at hand with a suitable (family of) “limiting problem(s)” where the symmetry is acting. We start with a simple example.

1. Pohoˇ zaev’s Non-existence Result Let Ω be a domain in IRn , n > 2. Consider the limit case p = 2∗ = Theorem I.2.1. Given λ ∈ IR, we would like to solve the problem −Δu = λu + u|u|2 u>0

(1.1) (1.2) (1.3)

∗

−2

u=0

2n n−2

in

in Ω , in Ω , on ∂Ω .

(Note that in order to be consistent with the literature, in this section we reverse the sign of λ as compared with Section I.2.1 or Section II.5.8.) As in Theorem I.2.1 we can approach this problem by a direct method and attempt to obtain non-trivial solutions of (1.1), (1.3) as relative minima of the functional 1 Iλ (u) = |∇u|2 − λ|u|2 dx , 2 Ω ∗

on the unit sphere in L2 (Ω), ∗

M = {u ∈ H01,2 (Ω) ; u2L2∗ = 1} . Equivalently, we may seek to minimize the Sobolev quotient

|∇u|2 − λ|u|2 dx 1,2 Ω Sλ (u; Ω) = 2/2∗ , u ∈ H0 (Ω) \ {0}. ∗ 2 |u| dx Ω Note that for λ = 0, as in Section I.4.4, S(Ω) =

inf 1,2

u∈H (Ω) 0 u=0

S0 (u; Ω) =

inf 1,2

u∈H (Ω) 0 u=0

|∇u|2 dx 2/2∗ |u|2∗ dx Ω Ω

is related to the (best) Lipschitz constant for the Sobolev embedding H01,2 (Ω) → ∗ L2 (Ω). Recall that for any u ∈ H01,2 (Ω) ⊂ D1,2 (IRn ) the ratio S0 (u; IRn ) is invariant under scaling u → uR (x) = u(Rx); that is, we have (1.4)

S0 (u; IRn ) = S0 (uR ; IRn ),

for all R > 0 .

In particular, as we already observed in Remark I.4.5, we have:

1. Pohoˇ zaev’s Non-existence Result

171

1.1 Lemma. S(Ω) = S is independent of Ω. Recalling Remark I.4.7, we note that Lemma 1.1 implies: 1.2 Theorem. S is never attained on a domain Ω ⊆ IRn , Ω = IRn . Hence, for λ = 0, the proof of Theorem I.2.1 necessarily fails in the limit case p = 2∗ . More generally, we have the following uniqueness result, due to Pohoˇzaev [1]: 1.3 Theorem. Suppose Ω = IRn is a smooth (possibly unbounded) domain in IRn , n ≥ 3, which is strictly star-shaped with respect to the origin in IRn , and let λ ≤ 0. Then any solution u ∈ H01,2 (Ω) of the boundary value problem (1.1), (1.3) vanishes identically. Theorem 1.3 applies to any solution, whereas Theorem 1.2 is limited to minima of S0 (·; Ω). However, Theorem 1.2 applies to any domain. The proof of Theorem 1.3 is based on the following “Pohoˇzaev identity”: 1.4 Lemma. Let g: IR → IR be continuous with primitive G(u) = and let u ∈ C 2 (Ω) ∩ C 1 (Ω) be a solution of the equation (1.5)

−Δu = g(u)

(1.6)

u=0

u 0

g(v) dv

in Ω on ∂Ω

in a domain Ω ⊂⊂ IRn . Then there holds 2 ∂u n−2 1 x · ν do = 0 , |∇u|2 dx − n G(u) dx + 2 2 ∂Ω ∂ν Ω Ω where ν denotes the exterior unit normal. Proof of Theorem 1.3. Let g(u) = λu + u|u|2 G(u) =

∗

−2

with primitive

∗ 1 λ 2 |u| + ∗ |u|2 . 2 2

By Theorems B.1 and B.2 and Lemma B.3 of Appendix B, any solution of (1.1), (1.3) is smooth on Ω. Hence from Pohoˇzaev’s identity we infer that 2 ∂u 1 2 ∗ x · ν do |∇u| dx − 2 G(u) dx + n − 2 ∂Ω ∂ν Ω Ω n|λ| 2 2∗ |∇u| − |u| dx + |u|2 dx = n−2 Ω Ω 2 ∂u 1 x · ν do = 0 . + ∂ν n−2 ∂Ω

172

Chapter III. Limit Cases of the Palais-Smale Condition

However, testing the Equation (1.1) with u, we infer that ∗ |∇u|2 − λ|u|2 − |u|2 dx = 0 , Ω

whence

Ω

2 ∂u x · ν do = 0 . ∂Ω ∂ν

|u|2 dx +

2|λ|

Moreover, since Ω is strictly star-shaped with respect to 0 ∈ IRn , we have = 0 on ∂Ω, and hence u ≡ 0 by the principle x · ν > 0 for all x ∈ ∂Ω. Thus ∂u ∂ν of unique continuation; see Heinz [2]. Proof of Lemma 1.4. Multiply (1.5) by x · ∇u and compute 0 = Δu + g(u) (x · ∇u) |∇u|2 + x · ∇G(u) = div ∇u(x · ∇u) − |∇u|2 − x · ∇ 2 n−2 |∇u|2 = div ∇u(x · ∇u) − x + xG(u) + |∇u|2 − nG(u) . 2 2 Upon integrating this identity over Ω and taking account of the fact that by (1.6) we have ∂u x · ∇u = x · ν on ∂Ω , ∂ν we obtain the asserted identity. d 1.5 Interpretation. Note that the function x · ∇u = dR uR used in the proof of Lemma 1.4 is the generator of the family of scaled maps {uR ; 0 < R < ∞}. This observation allows a connection between Theorem 1.3 and scale invariance of S = S0 . Indeed, we may interpret Theorem 1.3 as reﬂecting the non-compactness of the multiplicative group IR+ = {R ; 0 < R < ∞} acting on S via scaling. Note that this group action is a “manifest” symmetry of Sλ (· ; Ω) only in the case λ = 0 and Ω = IRn . In the case of a bounded domain Ω not all scalings u → uR will map H01,2 (Ω) into itself. For instance, if Ω is an annular region Ω = {x ; a < |x| < b}, in fact, H01,2 (Ω) does not admit any of these scalings as symmetries. (In Section 3 we will see that in this case (1.1)–(1.3) does have non-trivial solutions.) However, if Ω is star-shaped with respect to the origin, all scalings u → uR , R ≥ 1 will be symmetries of H01,2 (Ω), and compactness is lost as R → ∞. The eﬀect is shown in Theorem 1.3. Note that it is also possible to characterize solutions u ∈ H01,2 (Ω) of equation (1.1) as critical points of a functional Eλ on H01,2 (Ω) given by ∗ 1 1 |∇u|2 − λ|u|2 dx − ∗ |u|2 dx . (1.7) Eλ (u) = 2 Ω 2 Ω

2. The Brezis-Nirenberg Result

173

∗

By continuity of the embedding H01,2 (Ω) → L2 (Ω) → L2 (Ω), the functional Eλ is Fr´echet-diﬀerentiable on H01,2 (Ω). Moreover, for λ < λ1 , the ﬁrst Dirichlet eigenvalue of the operator −Δ, Eλ satisﬁes the conditions (1◦ )–(3◦ ) of the mountain pass lemma Theorem II.6.1; compare the proof of Theorem I.2.1. In view of Theorem II.6.1, the absence of a critical point u = 0 of Eλ for any λ ≤ 0 proves that Eλ for such λ cannot satisfy the Palais-Smale condition (P.-S.) on a star-shaped domain. Again the non-compact action R → uR (x) = u(Rx) can be held responsible.

2. The Brezis-Nirenberg Result In contrast to Theorem 1.3, for λ > 0 problem (1.1)–(1.3) may admit non-trivial solutions. However, a subtle dependence on the dimension n is observed. The ﬁrst result in this direction is due to Brezis and Nirenberg [2], building on ideas of Trudinger [1] and Aubin [2]. 2.1 Theorem. Suppose Ω is a domain in IRn , n ≥ 3, and let λ1 > 0 denote the ﬁrst eigenvalue of the operator −Δ with homogeneous Dirichlet boundary conditions. (1◦) If n ≥ 4, then for any λ ∈]0, λ1 [ there exists a solution of (1.1)–(1.3). (2◦) If n = 3, there exists λ∗ ∈ [0, λ1 [ such that for any λ ∈]λ∗ , λ1 [ problem (1.1)–(1.3) admits a solution. (3◦) If n = 3 and Ω = B1 (0) ⊂ IR3 , then λ∗ = λ41 and for λ ≤ λ41 there is no solution to (1.1)–(1.3). As we have seen in Section 1, there are (at least) two diﬀerent approaches to this theorem. The ﬁrst, which is the one chosen by Brezis and Nirenberg [2], involves the quotient

|∇u|2 − λ|u|2 dx Ω . Sλ (u; Ω) = 2/2∗ |u|2∗ dx Ω A second proof can be given along the lines of Theorem II.6.1, applied to the “free” functional Eλ ∗ 1 1 2 2 Eλ (u) = |u|2 dx |∇u| − λ|u| dx − ∗ 2 Ω 2 Ω deﬁned earlier. Recall that Eλ ∈ C 1 H01,2 (Ω) . As we shall see, while it is not true that Eλ satisﬁes the Palais-Smale condition “globally”, some compactness will hold in an energy range determined by the best Sobolev constant S; see Lemma 2.3 below. A similar compactness property holds for the functional Sλ . We will ﬁrst pursue the approach involving Sλ .

174

Chapter III. Limit Cases of the Palais-Smale Condition

Constrained Minimization Denote Sλ (Ω) =

inf

u∈H01,2 (Ω)\{0}

Sλ (u; Ω) .

Note that Sλ (Ω) ≤ S for all λ ≥ 0 (in fact, for all λ ∈ IR), and Sλ (Ω) in general is not attained. Similar to Theorem I.4.2 now there holds: 2.2 Lemma. If Ω is a bounded domain in IRn , n ≥ 3, and if Sλ (Ω) < S , then there exists u ∈

H01,2 (Ω),

u > 0, such that Sλ (Ω) = Sλ (u; Ω).

Proof. Consider a minimizing sequence (um ) for Sλ in H01,2 (Ω). Normalize um L2∗ = 1. Replacing um by |um |, if necessary, we may assume that um ≥ 0. Since by H¨ older’s inequality 2 2 |∇um |2 dx − c , |∇um | − λ|um | dx ≥ Sλ (um ; Ω) = Ω

Ω

H01,2 (Ω)

and strongly in L2 (Ω) as we also may assume that um u weakly in m → ∞. To proceed, observe that like (I.4.4) by Vitali’s convergence theorem we have ∗ ∗ |um |2 − |um − u|2 dx Ω

2∗ d um + (t − 1)u dt dx Ω 0 dt 1 2∗ −2 um + (t − 1)u um + (t − 1)u u dx dt = 2∗ 1

= (2.1)

0

→2

∗

Ω 1 0

t u|t u|2

∗

−2

Ω

∗

|u|2 dx as m → ∞ .

u dx dt = Ω

Also note that 2 2 |∇um | dx = |∇(um − u)| dx + |∇u|2 dx + o(1) , (2.2) Ω

Ω

Ω

where o(1) → 0 as m → ∞. Hence we obtain: |∇(um − u)|2 dx + |∇u|2 − λ|u|2 dx + o(1) Sλ (um ; Ω) + o(1) = Ω

≥ Sum −

Ω

u2L2∗ ∗

+

Sλ (Ω)u2L2∗ ∗

+ o(1)

≥ Sum − u2L2∗ + Sλ (Ω)u2L2∗ + o(1) ∗ ≥ S − Sλ (Ω) um − u2L2∗ + Sλ (Ω) + o(1) .

2. The Brezis-Nirenberg Result

175

Since S > Sλ (Ω) = limm→∞ Sλ (um ; Ω) by assumption, we ﬁnd that um → u in ∗ L2 (Ω), and uL2∗ = 1. By weak lower semi-continuity of the H01,2 (Ω)-norm then we have Sλ (u; Ω) ≤ lim Sλ (um ; Ω) = Sλ (Ω) , m→∞

as desired. Computing the ﬁrst variation of Sλ (u; Ω), as in the proof of Theorem I.2.1 we see that a positive multiple of u satisﬁes (1.1), (1.3). Since u ≥ 0, u = 0, from the strong maximum principle (Theorem B.4 of Appendix B) we infer that u > 0 in Ω. The proof is complete. The Unconstrained Case: Local Compactness Postponing the complete proof of Theorem 2.1 for a moment, we now also indicate the second approach, based on a careful study of the compactness properties of the free functional Eλ . Note that in the case of Theorem 2.1 both approaches are completely equivalent – and the ﬁnal step in the proof of Theorem 2.1 actually is identical in both cases. However, for more general nonlinearities with critical growth it is not always possible to reduce a boundary value problem like (II.6.1), (II.6.2) to a constrained minimization problem and we will have to use the free functional instead. Moreover, this second approach will bring out the peculiarities of the limiting case more clearly. Our presentation follows Cerami-Fortunato-Struwe [1]. An indication of Lemma 2.3 below is also given by Brezis-Nirenberg [2; p. 463]. 2.3 Lemma. Let Ω be a bounded domain in IRn , n ≥ 3, and let λ ∈ IR. Then any sequence (um ) in H01,2 (Ω) such that Eλ (um ) → β <

1 n/2 S , n

DEλ (um ) → 0 ,

as m → ∞, is relatively compact. Proof. To show boundedness of (um ), compute 2 o(1) 1 + um H 1,2 + S n/2 ≥ 2Eλ (um ) − um , DEλ (um ) 0 n 2∗ /2 2 2∗ 2 = 1− ∗ |um | dx ≥ c |um | dx , 2 Ω Ω where c > 0 and o(1) → 0 as m → ∞. Hence ∗ 2 2 2 |um | dx + ∗ |um |2 dx um H 1,2 = 2Eλ (um ) + λ 0 2 Ω Ω ≤ C + o(1)um H 1,2 , 0

and it follows that (um ) is bounded.

176

Chapter III. Limit Cases of the Palais-Smale Condition

Hence we may assume that um u weakly in H01,2 (Ω), and therefore also strongly in Lp (Ω) for all p < 2∗ by the Rellich-Kondrakov theorem; see Theorem A.5 of Appendix A. In particular, for any ϕ ∈ C0∞ (Ω) we obtain that ∗ ϕ, DEλ (um ) = ∇um ∇ϕ − λum ϕ − um |um |2 −2 ϕ dx Ω ∗ ∇u∇ϕ − λuϕ − u|u|2 −2 ϕ dx = ϕ, DEλ (u) = 0 , → Ω

as m → ∞. Hence, u ∈ H01,2 (Ω) weakly solves (1.1). Moreover, choosing ϕ = u, we have ∗ |∇u|2 − λ|u|2 − |u|2 dx , 0 = u, DEλ (u) = Ω

and hence

Eλ (u) =

1 1 − ∗ 2 2

∗

|u|2 dx = Ω

1 n

∗

|u|2 dx ≥ 0 . Ω

To proceed, note that by (2.1) and (2.2) we have 2 2 |∇um | dx = |∇(um − u)| dx + |∇u|2 dx + o(1) , Ω Ω Ω ∗ ∗ ∗ |um |2 dx = |um − u|2 dx + |u|2 dx + o(1) , Ω

Ω

Ω

where o(1) → 0 (m → ∞). Hence Eλ (um ) = Eλ (u) + E0 (um − u) + o(1). Similarly, again using (2.1) we have ∗ um |um |2 −2 (um − u) dx Ω ∗ 2∗ 2∗ |um | − |u| dx + o(1) = |um − u|2 dx + o(1) , = Ω

Ω

whence o(1) = um − u, DEλ (um ) = um − u, DEλ (um ) − DEλ (u) ∗ |∇(um − u)|2 − |um − u|2 dx + o(1) . = Ω

In particular, from the last equation it follows that 1 E0 (um − u) = |∇(um − u)|2 dx + o(1). n Ω On the other hand we have the bound

2. The Brezis-Nirenberg Result

E0 (um − u) = Eλ (um ) − Eλ (u) + o(1) 1 ≤ Eλ (um ) + o(1) ≤ c < S n/2 n

177

for m ≥ m0 .

Therefore um − u2H 1,2 ≤ c < S n/2 0

for m ≥ m0 .

But then Sobolev’s inequality um − u2L2∗ ≤ S −1 um − u2H 1,2 yields the 0 estimate ∗ ∗ −2 um − u2H 1,2 1 − S −2 /2 um − u2H 1,2 0 0 ∗ 2 |∇(um − u)| − |um − u|2 dx = o(1) , ≤ Ω

showing that um → u strongly in H01,2 (Ω), as desired. Lemma 2.3 motivates one to introduce the following variant of (P.-S.), which seems to appear ﬁrst in Brezis-Coron-Nirenberg [1]. 2.4 Definition. Let V be a Banach space, E ∈ C 1 (V ), β ∈ IR. We say that E satisﬁes condition (P.-S.)β , if any sequence (um ) in V such that E(um ) → β while DE(um ) → 0 as m → ∞ is relatively compact. (Such sequences in the sequel, for brevity, will be referred to as (P.-S.)β -sequences.) In terms of this deﬁnition, Lemma 2.3 simply says that the functional Eλ satisﬁes (P.-S.)β for any β < n1 S n/2 . Now recall that Eλ for λ < λ1 satisﬁes conditions (1◦ )–(3◦ ) of Theorem II.6.1. By Lemma 2.3, therefore, the proof of the ﬁrst two parts of Theorem 2.1 will be complete if we can show that for λ > 0 (respectively λ > λ∗ ) there holds (2.3)

β = inf sup Eλ (u) < p∈P u∈p

1 n/2 S , n

where, for a suitable function u1 satisfying Eλ (u1 ) ≤ 0, we let P = {p ∈ C 0 [0, 1] ; H01,2 (Ω) ; p(0) = 0, p(1) = u1 } , as in Theorem II.6.1. Of course, (2.3) and the condition Sλ (Ω) < S of Lemma 2.2 are related. Given u ∈ H01,2 (Ω) with uL2∗ = 1, we may let p(t) = t u, u1 = t1 u for suﬃciently large t1 to obtain 2 ∗ t t2 1 n/2 Sλ (u; Ω) − ∗ = Sλ (u; Ω) . β ≤ sup Eλ (t u) = sup 2 2 n 0≤t<∞ 0≤t<∞ Hence β ≤

1 n/2 n Sλ (Ω).

Likewise, any p ∈ P contains some u = 0 such that

178

Chapter III. Limit Cases of the Palais-Smale Condition

∗ |∇u|2 − λ|u|2 − |u|2 dx = 0 .

u, DEλ (u) = Ω

Indeed, since λ < λ1 , for u = p(t) = 0 with t close to 0 we have u, DEλ (u) > 0, while for u1 = p(1) we have u1 , DEλ (u1 ) < 2Eλ (u1 ) ≤ 0 , and by the intermediate value theorem there exists u, as claimed. But for such u we easily compute 1−2/2∗ 2 2 |∇u| − λ|u| dx Sλ (Ω) ≤ Sλ (u; Ω) = Ω 2/n

= (n Eλ (u))

≤

2/n n sup Eλ (u) . u∈p

Upon passing to the inﬁmum with respect to p ∈ P we ﬁnd that (2.4)

1 n/2 1 n/2 S (Ω) ≤ β = inf sup Eλ (u) ≤ Sλ (Ω) ; p∈P u∈p n λ n

that is, (2.3) and the condition Sλ < S are in fact equivalent. Proof of Theorem 2.1(1◦ ). It suﬃces to show that Sλ < S. Consider the family (2.5)

u∗ε (x)

=

[n(n − 2)ε2 ] [ε2 + |x|2 ]

n−2 4

n−2 2

, ε>0,

of functions u∗ε ∈ D1,2 (IRn ). Note that u∗ε (x) = ε the equation (2.6)

−Δu∗ε = u∗ε |u∗ε |2

∗

−2

2−n 2

u∗1

x ε

, and u∗ε satisﬁes

in IRn ,

as is easily veriﬁed by a direct computation. We claim that S0 u∗ε ; IRn = S; that is, the best Sobolev constant is achieved by the family u∗ε , ε > 0. Indeed, let u ∈ D1,2 (IRn ) satisfy S0 (u; IRn ) = S. (The existence of such a function u can be deduced for instance from Theorem I.4.9.) Using Schwarz-symmetrization we may assume that u is radially symmetric; that is, u(x) = u(|x|). (In fact, any positive solution of (2.6), which decays to 0 suﬃciently rapidly as |x| → ∞, by the result of Gidas-Ni-Nirenberg quoted in Section I.2.2 above is radially symmetric.) Moreover, u solves (2.6). Choose ε0 > 0 such that u∗ε0 (0) = u(0). Then u and u∗ε0 both are solutions of the ordinary diﬀerential equation of second order in r = |x|, ∗ ∂ ∂ for r > 0 , r 1−n r n−1 u = u|u|2 −2 ∂r ∂r sharing the initial data u(0) = u∗ε0 (0), ∂r u(0) = ∂r u∗ε0 (0) = 0. It is not hard to prove that this initial value problem admits a unique solution. Thus u = u∗ε0 , which implies that S = S0 (u; IRn ) = S0 (u∗ε0 ; IRn ) = S0 (u∗ε ; IRn ) for all ε > 0.

2. The Brezis-Nirenberg Result

179

∗

Since (2.6) also yields that u∗ε 2H 1,2 = u∗ε 2L2∗ it follows that 0

∗

u∗ε 2H 1,2 = u∗ε 2L2∗ = S n/2 for all ε > 0 . 0

We may suppose that 0 ∈ Ω. Let η ∈ C0∞ (Ω) be a ﬁxed cut-oﬀ function, η ≡ 1 in a neighborhood Bρ (0) of 0. Let uε = η u∗ε and compute |∇uε |2 dx = |∇u∗ε |2 η 2 dx + O εn−2 Ω Ω |∇u∗ε |2 dx + O εn−2 = S n/2 + O εn−2 . = n IR ∗ 2∗ |uε | dx = |u∗ε |2 dx + O εn = S n/2 + O εn n Ω IR |uε |2 dx ≥ |u∗ε |2 dx Ω

(2.7)

Bρ (0)

≥

Bε (0)

[n(n − 2)ε2 ] [2ε2 ]n−2

n−2 2

dx n−2

[n(n − 2)ε2 ] 2 [2|x|2 ]n−2 Bρ (0)\Bε (0) ρ = c1 · ε2 + c2 εn−2 r 3−n dr n−2 ε ⎧ 2 ⎨ c ε + O ε , if n > 4 2 = c ε2 | ln ε| O ε , if n = 4 + ⎩ 2 if n = 3 cε+ O ε , +

dx

with positive constants c, c1 , c2 > 0. Thus, if n ≥ 5 (S n/2 − cλε2 + O(εn−2 )) (S n/2 + O(εn ))2/2∗ = S − cλε2 + O εn−2 < S ,

Sλ (uε ; Ω) ≤

if ε > 0 is suﬃciently small. Similarly, if n = 4, we have Sλ (uε ; Ω) ≤ S − cλε2 | ln ε| + O ε2 < S for ε > 0 suﬃciently small. Remark on Theorem 2.1(2◦ ),(3◦ ). If n = 3, estimate (2.7) shows that the “gain” due to the presence of λ and the “loss” due to truncation of u∗ε may be of the same order in ε; hence Sλ can only be expected to be smaller than S for “large” λ. To see that λ∗ < λ1 , choose the ﬁrst eigenfunction u = ϕ1 of (−Δ) as comparison function. The non-existence result for Ω = B1 (0), λ ≤ λ41 follows from a weighted estimate similar to Lemma 1.4; see Brezis-Nirenberg [2; Lemma 1.4]. We omit the details.

180

Chapter III. Limit Cases of the Palais-Smale Condition

Theorem 2.1 should be viewed together with the global bifurcation result of Rabinowitz [1; p. 195 f.]. Intuitively, Theorem 2.1 indicates that the branch of positive solutions found by Rabinowitz in dimension n ≥ 4 on a star-shaped domain bends back to λ = 0 and becomes asymptotic to this axis.

Fig. 2.1. Solution “branches” for (1.1), (1.3) depending on λ

However, Equation (1.1) may have many positive solutions. In Section 3 we shall see that if Ω is an annulus, positive radial solutions exist for any value of λ < λ1 . But note that, by Theorem 1.2, for λ = 0 these cannot minimize Sλ . Hence we may have many diﬀerent branches of (or secondary bifurcations of branches of) positive solutions, in general. Multiple Solutions 2.5 Bifurcation from higher eigenvalue. If we drop the requirement of positivity, the local Palais-Smale condition Lemma 2.3 permits us to obtain bifurcation of non-trivial solutions of (1.1), (1.3) from higher eigenvalues, as well. Recall that by a result of B¨ ohme [1] and Marino [1] it is known that any eigenvalue of −Δ on H01,2 (Ω) is a point of bifurcation from the trivial solution of (1.1), (1.3). However, the variational method may give better estimates for the λ-interval of existence for such solutions. The following result is due to Cerami-Fortunato-Struwe [1]: 2.6 Theorem. Let Ω be a bounded domain in IRn , n ≥ 3, and let 0 < λ1 < λ2 ≤ . . . denote the eigenvalues of −Δ in H01,2 (Ω). Also let −2/n >0. ν = S · Ln (Ω) Then, if

2. The Brezis-Nirenberg Result

181

m = m(λ) = #{j ; λ < λj < λ + ν} , problem (1.1), (1.3) admits at least m distinct pairs of non-trivial solutions. Remark. From the Weyl formula λj ∼ C(Ω)j 2/n for the asymptotic behavior of the eigenvalues λj we conclude that m(λ) → ∞ as λ → ∞. Proof of Theorem 2.6. Consider the functional

|∇u|2 − λ|u|2 dx Ω Sλ (u) := Sλ (u; Ω) = 2/2∗ |u|2∗ dx Ω on the unit sphere M = {u ∈ H01,2 (Ω) ; uL2∗ = 1} . M ⊂ H01,2 (Ω) is a complete Hilbert manifold, invariant under the involution u → −u. Recall that Sλ is diﬀerentiable on M1 and that if u ∈ M is a critical ˜ = β 2∗ −2 u solves (1.1), (1.3) with point of Sλ with Sλ (u) = β > 0, then u u) = Eλ (˜

1 Sλ (u)n/2 ; n

see the proof of Theorem I.2.1. Moreover, (um ) is a (P.-S.)β -sequence for Sλ if and only if the sequence 1 ˜m = β 2∗ −2 um , is a (P.-S.)β˜ sequence for Eλ with β˜ = n1 β n/2 . (˜ um ), where u In particular, by Lemma 2.3 we have that Sλ satisﬁes (P.-S.)β on M for any β ∈]0, S[. Now let γ denote the Krasnoselskii genus, and for j ∈ IN such that λj ∈ ]λ, λ + ν[ let βj = inf sup Sλ (u) . A⊂M γ(A)≥j

u∈A

Note that by Proposition II.5.3 for any A ⊂ M such that γ(A) ≥ j there exist at least j mutually orthogonal vectors in A, whence βj > 0 for all j as above. Moreover, if we denote by ϕk ∈ H01,2 (Ω) the kth eigenfunction of −Δ and let Aj = span{ϕ1 , . . . , ϕj } ∩ M , we obtain that

|u|2 dx βj ≤ sup Sλ (u) ≤ (λj − λ) sup Ω 2/2∗ u∈Aj u∈Aj |u|2∗ dx Ω 2 < ν Ln (Ω) n = S . Hence the theorem follows from Theorem II.4.2 and Lemma II.5.6. We do not know how far the solution branches bifurcating oﬀ the trivial solution u ≡ 0 at λ = λj extend. However, the following result has been obtained by Capozzi-Fortunato-Palmieri [1]:

182

Chapter III. Limit Cases of the Palais-Smale Condition

2.7 Theorem. Let Ω be a bounded domain in IRn , n ≥ 4. Then for any λ > 0 problem (1.1), (1.3) admits a non-trivial solution. The proof of Capozzi-Fortunato-Palmieri is based on a linking argument and the local Palais-Smale condition Lemma 2.3. A simpler proof for λ not belonging to the spectrum of −Δ, using the duality method, was worked out by Ambrosetti and Struwe [1]. Theorem 2.7 leaves open the question of multiplicity. For n ≥ 4, λ > 0 Fortunato-Jannelli [1] present results in this direction using symmetries of the domain to restrict the space of admissible functions to certain symmetric subspaces, where the ﬁrst Dirichlet eigenvalue is increased suﬃciently for Theorem 2.1 to become applicable. For instance, if Ω = B1 (0; IR2 ) × Ω ⊂ IRn , n ≥ 4, λ > 0, essentially Fortunato-Jannelli solve (1.1), (1.3) by constructing a positive solution u to (1.1), (1.3) on a “slice” Ωm = {x = reiϕ ∈ B1 (0; IR2 ) ; 0 ≤ ϕ ≤

π } × Ω m

and reﬂecting u in the “vertical” edges of Ωm a total of 2m times. Since the (m) ﬁrst eigenvalue λ1 of (−Δ), acting on H01,2 (Ωm ), tends to ∞ as m → ∞, (m) we can achieve that λ1 > λ for m ≥ m0 . Hence the existence of a positive solution u to (1.1), (1.3) on Ωm is guaranteed by Theorem 2.1.

Fig. 2.2. A slice Ωm of the pie Ω

A diﬀerent idea is used by Ding [1] to construct inﬁnitely many solutions of varying sign of Equation (2.6) on IRn . In fact, Ding’s result involves a beautiful combination of analysis and geometry. Conformally mapping IRn to the nsphere and imposing invariance with respect to a subgroup SO(k) ⊂ SO(n) of rotations of the sphere as a constraint on admissible comparison functions,

3. The Eﬀect of Topology

183

Ding eﬀectively reduces Equation (2.6) to a sub-critical variational problem to which Theorem 4.2 may be applied. For the existence of multiple radially symmetric solutions on balls Ω = BR (0) ⊂ IRn a subtle dependence on the space dimension is observed, related to the fact that the best Sobolev constant S is attained in the class of radially symmetric functions (on IRn ). The following result is due to Cerami-SoliminiStruwe [1] and Solimini [1] : 2.8 Theorem. Suppose Ω = BR (0) is a ball in IRn , n ≥ 7. Then for any λ > 0 problem (1.1), (1.3) admits inﬁnitely many radially symmetric solutions. The proof of Theorem 2.8 uses a characterization of diﬀerent solutions by the nodal properties they possess, as in the locally compact case; see Remark II.7.3. By recent results of Atkinson-Brezis-Peletier [1] the restriction on the dimension n in Theorem 2.8 appears to be sharp; see also Adimurthi-Yadava [1]. In 2002, ﬁnally, Devillanova-Solimini [1] showed that for any λ > 0 problem (1.1), (1.3) admits inﬁnitely many solutions also on an arbitrary smoothly bounded domain Ω ⊂ IRn , provided that n ≥ 7. Their method of proof relies on a precise analysis as p ↑ 2∗ of solutions u = up to the sub-critical equations −Δu = λu + u|u|p−2 in Ω , with u = 0 on ∂Ω. Again their method indicates that the restriction n ≥ 7 may be sharp. 2.9 Notes. (1◦ ) Egnell [1], Guedda-Veron [1], and Lions-Pacella-Tricarico [1] have studied problems of the type (1.1), (1.3) involving the (degenerate) pseudo-Laplace operator and partially free boundary conditions. (2◦) The linear term λu in Equation (1.1) may be replaced by other compact perturbations; see Brezis-Nirenberg [2]. In this regard we also mention the results by Mancini-Musina [1] concerning obstacle problems of type (1.1), (1.3). Still a diﬀerent kind of “perturbation” will be considered in the next section.

3. The Eﬀect of Topology Instead of a star-shaped domain, as in Theorem 1.3, consider an annulus Ω = {x ∈ IRn ; r1 < |x| < r2 } , and for λ ∈ IR let Eλ be given by (1.7). Note that if we restrict our attention to radial functions 1,2 H0,rad (Ω) = {u ∈ H01,2 (Ω) ; u(x) = u(|x|)} ,

by estimate (II.11.6) the embedding

184

Chapter III. Limit Cases of the Palais-Smale Condition 1,2 H0,rad (Ω) → Lp (Ω)

1,2 is compact for any p < ∞. Hence DEλ : H0,rad (Ω) → H −1 (Ω) is of the form id + compact. Since any (P.-S.)-sequence for Eλ by the proof of Lemma 2.3 is bounded, from Proposition II.2.2 we thus infer that Eλ satisﬁes (P.-S.) globally 1,2 (Ω) and hence from Theorem II.5.7 or Theorem II.6.5 that problem on H0,rad (1.1), (1.3) possesses inﬁnitely many radially symmetric solutions on Ω, for any n ≥ 2, and any λ ∈ IR, in particular, for λ = 0. This result stands in striking contrast with Theorem 1.3 (or Theorem 2.8, as regards the restriction of the dimension). In the following we shall investigate whether the solvability of (1.1), (1.3) on an annulus is a singular phenomenon, observable only in a highly symmetric case, or is stable and survives perturbations of the domain.

A Global Compactness Result Remark that by Theorem 1.2, for λ = 0 no non-trivial solution u ∈ H01,2 (Ω) of (1.1) can satisfy Sλ (u; Ω) ≤ S. Hence the local compactness of Lemma 2.3 will not suﬃce to produce such solutions and we must study the compactness properties of Eλ , respectively Sλ , at higher energy levels as well. The next result can be viewed as an extension of P.-L. Lions’ concentration-compactness principle for minimization problems (see Section I.4) to problems of minimax type. The idea of analyzing the behavior of a (P.-S.)-sequence near points of concentration by “blowing up” the singularities seems to appear ﬁrst in papers by Sacks-Uhlenbeck [1] and Wente [5], where variants of the local compactness condition Lemma 2.3 are obtained (see Sacks-Uhlenbeck [1; Lemma 4.2]). In the next result, due to Struwe [8], we systematically employ the blow-up technique to characterize all energy values β of a variational problem where (P.-S.)β may fail. 3.1 Theorem. Suppose Ω is a smoothly bounded domain in IRn , n ≥ 3, and for λ ∈ IR let (um ) be a (P.-S.)-sequence for Eλ in H01,2 (Ω) ⊂ D1,2 (IRn ). Then j ), (xjm ), 1 ≤ j ≤ k, of radii there exist a number k ∈ IN0 , sequences (Rm j → ∞ (m → ∞) and points xjm ∈ Ω, a solution u0 ∈ H01,2 (Ω) ⊂ D1,2 (IRn ) Rm to (1.1), (1.3) and non-trivial solutions uj ∈ D1,2 (IRn ), 1 ≤ j ≤ k, to the “limiting problem” associated with (1.1) and (1.3), (3.1)

−Δu = u|u|2

∗

−2

in IRn ,

such that a subsequence (um ) satisﬁes % % k % % j % %um − u0 − u m% % j=1

Here ujm denotes the rescaled function

D1,2 (IRn )

→0.

3. The Eﬀect of Topology j ujm (x) = (Rm )

n−2 2

185

j uj R m (x − xjm ) , 1 ≤ j ≤ k, m ∈ IN .

Moreover, we have Eλ (um ) → Eλ (u0 ) +

k

E0 (uj ) .

j=1

1,2 3.2 Remarks. (1◦ ) In particular, if Ω is a ball Ω = BR (0), um ∈ H0,rad (Ω), ∗ from the uniqueness of the family (uε )ε>0 of radial solutions to (3.1) – see the proof of Theorem 2.1(1◦ ) – it follows that each uj is of the form (2.5) with E0 (uj ) = n1 S n/2 =: β ∗ . Hence in this case (P.-S.)β holds for Eλ for all levels β which cannot be decomposed

β = β0 + kβ ∗ , where k ≥ 1 and β0 = Eλ (u0 ) is the energy of some radial solution of (1.1), (1.3). Similarly, if Ω is an arbitrary bounded domain and um ≥ 0 for all m, then also uj ≥ 0 for all j, and by a result of Gidas-Ni-Nirenberg [1; p. 210 f.] and Obata [1] again each function uj will be radially symmetric about some point xj . Therefore also in this case each uj is of the form uj = u∗ε (· − xj ) for some ε > 0, and (P.-S.)β holds for all β which are not of the form β = β0 + kβ ∗ where k ≥ 1 and β0 = Eλ (u0 ) is the energy of some non-negative solution u0 of (1.1), (1.3). (2◦) For some time it was believed that the family (2.5) gives all non-trivial solutions of (3.1). However, Ding’s [1] result shows that (3.1) also admits inﬁnitely many solutions of changing sign which are distinct modulo scaling. (3◦) In general, decomposing a solution v of (3.1) into positive and negative parts v = v+ + v− , where v± = ± max{±v, 0}, upon testing (3.1) with v± from Sobolev’s inequality we infer that ∗ −Δv − v|v|2 −2 v± dx 0= n IR ∗ ∗ ∗ −2 |∇v± |2 − |v± |2 dx ≥ 1 − S −2 /2 v± 2D1,2 v± 2D1,2 . = IRn

Hence, either v ≡ 0, or E0 (v± ) =

1 1 v± 2D1,2 ≥ S n/2 = β ∗ . n n

In fact, E0 (v± ) > β ∗ ; otherwise S would be achieved at v± , contradicting Theorem 1.2. Therefore any solution v of (3.1) that changes sign satisﬁes E0 (v) = E0 (v+ ) + E0 (v− ) > 2β ∗ , and in Theorem 3.1 we can assert that E0 (uj ) ∈ {β ∗ } ∪ ]2β ∗ , ∞[.

186

Chapter III. Limit Cases of the Palais-Smale Condition

In particular, if (1.1), (1.3) does not admit any solution but the trivial solution u ≡ 0, the local Palais-Smale condition (P.-S.)β will hold for all β < 2β ∗ , except for β = β ∗ . When λ = 0, as noted above, any non-trivial solution u of problem (1.1), (1.3) on a bounded domain satisﬁes E0 (u) > β ∗ . Theorem 3.1 then likewise guarantees that condition (P.-S.)β will hold for all β < 2β ∗ , except for β = β ∗ . See Weth [1] for a recent improvement of this result. (4◦) Bahri-Coron [1] observe that in Theorem 3.1 in addition we can assert i Rm j Rm

+

j Rm i j + Rm Rm |xim − xjm |2 → ∞ as m → ∞ for all i = j. i Rm

Proof of Theorem 3.1. First recall that as in the proof of Lemma 2.3 any (P.S.)-sequence for Eλ is bounded. Hence we may assume that um u0 weakly in H01,2 (Ω), and u0 solves (1.1), (1.3). Moreover, if we let vm = um − u0 we have vm → 0 strongly in L2 (Ω), and by (2.1), (2.2) also that ∗ 2∗ 2∗ |vm | dx = |um | dx − |u0 |2 dx + o(1) , Ω Ω Ω 2 2 |∇vm | dx = |∇um | dx − |∇u0 |2 dx + o(1) , Ω

Ω

Ω

where o(1) → 0 (m → ∞). Hence, in particular, we obtain that Eλ (um ) = Eλ (u0 ) + E0 (vm ) + o(1) . Also note that DEλ (um ) = DEλ (u0 ) + DE0 (vm ) + o(1) = DE0 (vm ) + o(1) , where o(1) → 0 in H −1 (Ω) (m → ∞). Using the following lemma, we can now proceed by induction: 3.3 Lemma. Suppose (vm ) is a (P.-S.)-sequence for E = E0 in H01,2 (Ω) such that vm 0 weakly. Then there exists a sequence (xm ) of points xm ∈ Ω, a sequence (Rm ) of radii Rm → ∞ (m → ∞), a non-trivial solution v 0 to the limiting problem (3.1) and a (P.-S.)-sequence (wm ) for E0 in H01,2 (Ω) such that for a subsequence (vm ) there holds n−2 wm = vm − Rm2 v 0 Rm (· − xm ) + o(1) ,

where o(1) → 0 in D1,2 (IRn ) as m → ∞. In particular, wm 0 weakly. Furthermore, E0 (wm ) = E0 (vm ) − E0 (v 0 ) + o(1) . Moreover, Rm dist(xm , ∂Ω) → ∞ .

3. The Eﬀect of Topology

187

Finally, if E0 (vm ) → β < β ∗ , the sequence (vm ) is relatively compact and hence vm → 0 strongly in H01,2 (Ω), E0 (vm ) → β = 0 as m → ∞. 1 Proof of Theorem 3.1 (completed). Apply Lemma 3.3 to the sequences vm = j−1 i 0 j 0 j−1 j−1 um − u , vm = um − u − i=1 um = vm − um , j > 1, where i i n−2 uim (x) = (Rm ) 2 ui R m (x − xim ) .

By induction j E0 (vm ) = Eλ (um ) − Eλ (u0 ) −

j−1

E0 (ui )

i=1

≤ Eλ (um ) − (j − 1)β ∗ . Since the latter will be negative for large j, by Lemma 3.3 the induction will terminate after some index k ≥ 0. Moreover, for this index we have k+1 vm = um − u0 −

k

ujm → 0

j=1

strongly in D1,2 (IRn ), and Eλ (um ) − Eλ (u0 ) −

k

E0 (uj ) → 0 ,

j=1

as desired. Proof of Lemma 3.3. If E0 (vm ) → β < β ∗ , by Lemma 2.3 the sequence (vm ) is strongly relatively compact and hence vm → 0, β = 0. Therefore, we may assume that E0 (vm ) → β ≥ β ∗ = n1 S n/2 . Moreover, since DE0 (vm ) → 0 we also have 1 1 1 |∇vm |2 dx = E0 (vm ) − ∗ vm , DE0 (vm ) → β ≥ S n/2 n Ω 2 n and hence that (3.2)

|∇vm |2 dx = nβ ≥ S n/2 .

lim inf m→∞

Ω

Extend vm ≡ 0 outside Ω. Denote

|∇vm |2 dx

Qm (r) = sup x∈Ω

Br (x)

the concentration function of vm , introduced in Section I.4.3. Choose xm ∈ IRn and scale 2−n vm → v˜m (x) = Rm2 vm x/Rm + xm

188

Chapter III. Limit Cases of the Palais-Smale Condition

such that ˜ m (1) = sup Q

x∈IRn

|∇˜ vm |2 dx =

B1 (x)

|∇˜ vm |2 dx = B1 (0)

1 n/2 , S 2L

where L is a number such that B2 (0) is covered by L balls of radius 1. Clearly, by (3.2) we have Rm ≥ R0 > 0, uniformly in m. ˜ m = {x ∈ IRn ; x/Rm + xm ∈ Ω}, we may assume that Ω ˜m → Letting Ω n ˜ Ω∞ ⊂ IR . Since ˜ vm 2D1,2 = vm 2D1,2 → nβ < ∞, ˜∞ ) we also may assume that v˜m v 0 weakly in D1,2 (IRn ). By density of C0∞ (Ω 1,2 ˜ 1,2 ˜ 0 in D (Ω∞ ), moreover, we can ﬁnd functions v˜m ∈ H0 (Ωm ) such that v 0 = 0 ˜ ∞ ). in D1,2 (Ω limm→∞ v˜m We claim that v˜m → v 0 strongly in H 1,2 (Ω ), or, equivalently, that the 0 ) → 0 strongly in H 1,2 (Ω ) for any Ω ⊂⊂ IRn . It suﬃces functions (˜ vm − v˜m to consider balls Ω = B1 (x0 ) for any x0 ∈ IRn . For brevity in the following we let Br (x0 ) =: Br . Choose cut-oﬀ functions ϕi ∈ C0∞ (IRn ) such that 0 ≤ ϕi ≤ 1, i = 1, 2, ϕ1 ≡ 1 in B1 , ϕ1 ≡ 0 outside B3/2 , while ϕ2 ≡ 1 in B3/2 , ϕ2 ≡ 0 outside B2 . i 0 ˜m ), i = 1, 2. Note that we have = (˜ vm − v˜m )ϕi ∈ H01,2 (Ω Then let w ˜m i 0 H 1,2 (Ω˜m ) ≤ c˜ vm − v˜m H 1,2 (Ω˜m ) ≤ C < ∞, i = 1, 2. w ˜m 1 2 is supported in the ball B3/2 , and w ˜m is supported in B2 . Moreover, w ˜m n 0 1,2 Recall that (˜ vm − v˜m ) 0 weakly in D (IR ). The restrictions of these functions to B2 then converge weakly to 0 in H 1,2 (B2 ). From compactness of the embedding H 1,2 (B2 ) → Lp (B2 ) for any p < 2∗ , ﬁnally, we also ﬁnd that 0 ) → 0 strongly in Lp (B2 ) for any such p. (˜ vm − v˜m Using convergence arguments familiar by now and Sobolev’s inequality, we then obtain

(3.3)

1 ˜ m ) o(1) = w ˜m ϕ1 , DE0 (˜ vm ; Ω ∗ 1 1 ∇˜ vm ∇w ϕ1 dx + o(1) ˜m − v˜m |˜ vm |2 −2 w ˜m = n IR 0 2 0 2∗ )| − |˜ vm − v˜m | ϕ21 dx + o(1) |∇(˜ vm − v˜m = n IR 0 2 2 0 2∗ 2 2∗ −2 |∇(˜ vm − v˜m dx + o(1) = )| ϕ1 − |˜ vm − v˜m | ϕ1 ϕ2 n IR 1 2 1 2 2 2∗ −2 |∇w ˜m dx + o(1) = | − |w ˜m | |w ˜m | IRn ∗ 1 2 2 2∗ −2 ≥ w ˜m D1,2 (IRn ) 1 − S −2 /2 w ˜m D1,2 (IRn ) + o(1) ,

where o(1) → 0 as m → ∞. But note that

3. The Eﬀect of Topology

189

2 2 0 2 2 |∇w ˜m | dx = |∇(˜ vm − v˜m )| ϕ2 dx + o(1) IRn IRn 2 2 0 2 |∇˜ vm | − |∇v )| ϕ2 dx + o(1) ≤ |∇˜ vm |2 dx + o(1) =

IRn

B2

˜ m (1) + o(1) = 1 S n/2 + o(1) . ≤LQ 2 ∗

∗

2 2 −2 ˜m D1,2 (IRn ) ≤ c < 1 for large m, and from (3.3) it follows that Thus S −2 /2 w 1 1,2 w ˜m → 0 in D (IRn ); that is, v˜m → v 0 strongly locally in H 1,2 , as desired. In particular, 1 n/2 |∇v 0 |2 dx = >0, S 2L B1 (0)

and v 0 ≡ 0. Since the original sequence vm 0 weakly, thus it also follows ˜ ∞ either equals IRn or is a half-space. Now that Rm → ∞ as m → ∞, and Ω we distinguish two cases: ˜∞ is a half-space, or (1◦) Rm dist(xm , ∂Ω) ≤ c < ∞, uniformly, in which case Ω ◦ ˜ ˜ (2 ) Rm dist(xm , ∂Ω) → ∞, in which case Ωm → Ω∞ = IRn . ˜ ∞ ) we have that ϕ ∈ C ∞ (Ω ˜m ) for Since in each case for any ϕ ∈ C0∞ (Ω 0 large m, there holds ˜ ∞ ) = lim ϕ, DE0 (˜ ˜ m ) ≤ C lim ||DE0 (vm )||H −1 = 0 , ϕ, DE0 (v 0 ; Ω vm ; Ω m→∞

m→∞

˜ ∞ ) is a weak solution of (3.1) on Ω ˜∞ . But for all such ϕ, and v 0 ∈ H01,2 (Ω ˜ ∞ = IRn+ , by Theorem 1.3 then v 0 must vanish identically. Thus (1◦ ) is if Ω impossible, and we are left with (2◦ ). Thus we may now also choose a sequence ˜ m := Rm (Rm )−1 → ∞ while Rm dist(xm , ∂Ω) → ∞ as (Rm ) such that R m → ∞. To conclude the proof, again let ϕ ∈ C0∞ (IRn ) be a cut-oﬀ function satisfying 0 ≤ ϕ ≤ 1, ϕ ≡ 1 in B1 (0), ϕ ≡ 0 outside B2 (0), and let n−2 wm (x) = vm (x) − Rm2 v 0 Rm (x − xm ) · ϕ Rm (x − xm ) ∈ H01,2 (Ω).

That is, 2−n

˜ m) . w ˜m (x) = Rm2 wm (x/Rm + xm ) = v˜m (x) − v 0 (x)ϕ(x/R ˜ m . Note that Set ϕm (x) = ϕ x/R 0 ∇ v (ϕm − 1) 2 dx IRn 2 0 2 2 |∇v | (ϕm − 1) dx + C |v 0 |2 ∇(ϕm − 1) dx ≤C n n IR IR 0 2 −2 ˜ ≤C |∇v | dx + C R |v 0 |2 dx . m IRn \BR ˜ m (0)

B2R ˜ m (0)\BR ˜ m (0)

190

Chapter III. Limit Cases of the Palais-Smale Condition

But ∇v 0 ∈ L2 (IRn ). Therefore the ﬁrst term tends to 0 as m → ∞, while by H¨ older’s inequality also the second term ˜ −2 R m

2/2∗ 0 2∗

|v | dx ≤ C

|v |

0 2

B2R ˜ m (0)\BR ˜ m (0)

dx

→0

B2R ˜ m (0)\BR ˜ m (0)

as m → ∞. Thus we have w ˜m = v˜m − v 0 + o(1), where o(1) → 0 in D1,2 (IRn ). As in the proof of Lemma 2.3 we then obtain the expansion ˜m ) = E0 (˜ vm ) − E0 (v 0 ) + o(1) = E0 (vm ) − E0 (v 0 ) + o(1). E0 (wm ) = E0 (w Moreover, by a similar reasoning for any sequence of testing functions ψm ∈ ˜ m ) with ψm H 1,2 ≤ 1 with error o(1) → 0 as m → ∞ we obtain H 1,2 (Ω ∗ ˜ ψm , DE0 (w ˜m ; Ωm ) = ˜m | w ˜m |2 −2 ψm dx ∇w ˜m ∇ψm − w n IR ∗ ∇˜ vm ∇ψm − v˜m |˜ = vm |2 −2 ψm dx IRn 0 ∗ ∇v ∇ψm − v 0 |v 0 |2 −2 ψm dx + o(1) − IRn

˜ m ) − ψm , DE0 (v 0 ; IRn ) + o(1). = ψm , DE0 (˜ vm ; Ω Thus, we ﬁnd DE0 (wm ; Ω)H −1 ˜ m )H −1 = = DE0 (w ˜m ; Ω

sup ˜m ) ψ∈H 1,2 (Ω ψ 1,2 ≤1 H

˜ m ) ψ, DE0 (w ˜m ; Ω

˜ m )H −1 + DE0 (v 0 ; IRn )H −1 + o(1) vm ; Ω ≤ DE0 (˜ = DE0 (vm ; Ω)H −1 + o(1) → 0

(m → ∞) .

This concludes the proof. Positive Solutions on Annular-shaped Regions With the aid of Theorem 3.1 we can now show the existence of positive solutions to (1.1), (1.3) on perturbed annular domains for λ = 0. The following result is due to Coron [2]: 3.4 Theorem. Suppose Ω is a bounded domain in IRn satisfying the following condition: There exist constants 0 < R1 < R2 < ∞ such that (1◦ ) ◦

(2 )

Ω ⊃ {x ∈ IRn ; R1 < |x| < R2 } , Ω ⊃ {x ∈ IRn ; |x| < R1 } .

3. The Eﬀect of Topology

191

Then, if R2 /R1 is suﬃciently large, problem (1.1), (1.3) for λ = 0 admits a positive solution u ∈ H01,2 (Ω). Again note that the solution u must have an energy above the compactness threshold given by Lemma 2.3. The idea of the proof is to argue by contradiction and to use a minimax method for S = S0 ( · ; Ω). Recall that the Sobolev constant in IRn is attained at the functions u∗ε deﬁned in (2.5). Consider the sphere Σ = {x ∈ IRn ; |x|2 = R1 R2 } in the annulus {x ∈ IRn ; R1 < |x| < R2 }. For suitably small ε > 0 then consider the collection A = {u∗ε (· − σ)ϕ ∈ H01,2 (Ω); σ ∈ Σ} of shifted functions u∗ε , truncated with a suitable cut-oﬀ function ϕ ∈ C0∞ (Ω). If we regard the center of mass, as explained below in detail, then for suﬃciently small ε > 0 the set A is homeomorphic to the sphere Σ, which by assumption (2◦ ) is not contractible in Ω. After normalizing, A ⊂ M = {u ∈ H01,2 (Ω) : uL2∗ = 1}. On the other hand, any set A of positive functions in M is contractible in M . Moreover, for a suﬃciently large ratio R2 /R1 we can contract A in such a way that the corresponding minimax-value β is smaller than the number 22/n S. If we assume that (1.1), (1.3) does not admit a positive solution, by Theorem 3.1 and Remark 3.2 this is the smallest number above S where the Palais-Smale condition fails in M . By applying the deformation lemma Theorem II.3.11 we then conclude that the set A can be contracted below the level S + δ for any δ > 0. For suitably small δ > 0, ﬁnally, such a contraction will induce a contraction of Σ in Ω, and the desired contradiction will result. Proof. We may assume R1 = (4R)−1 < 1 < 4R = R2 . Consider the unit sphere Σ = {x ∈ IRn ; |x| = 1} . For σ ∈ Σ, x ∈ IRn , 0 ≤ t < 1 let # uσt (x) =

1−t (1 − t)2 + |x − tσ|2

$ n−2 2

∈ D1,2 (IRn ) .

Note that S is attained on any such function uσt , and uσt “concentrates” at σ as t → 1. Moreover, letting t → 0 we have #

uσt

1 → u0 = 1 + |x|2

$ n−2 2 ,

for any σ ∈ Σ. Choose a radially symmetric function ϕ ∈ C0∞ (IRn ) such that 0 ≤ ϕ ≤ 1 on Ω, ϕ ≡ 1 on the annulus {x ; 12 < |x| < 2} and ϕ ≡ 0 outside the annulus {x ; 14 < |x| < 4}. Given any R ≥ 1, scale ⎧ ⎨ ϕ(Rx), 0 ≤ |x| < R−1 ϕR (x) = 1, R−1 ≤ |x| < R ⎩ ϕ(x/R), R ≤ |x| and let

192

Chapter III. Limit Cases of the Palais-Smale Condition

wtσ = uσt · ϕR , w0 = u0 · ϕR ∈ H01,2 (Ω) . Note that |∇(wtσ − uσt )|2 dx ≤ c IRn

(IRn \B2R )∪B(2R)−1

+ c · R−2

|∇uσt |2 dx

B4R \B2R

|uσt |2 dx + cR2

|uσt |2 dx → 0 B(2R)−1

as R → ∞, uniformly in σ ∈ Σ, t ∈ [0, 1[. Also consider the normalized functions 9 9 vtσ = wtσ wtσ L2∗ , v0 = w0 w0 L2∗ . Since wtσ − uσt D1,2 → 0 as R → ∞ and uσt D1,2 = u0 D1,2 > 0, we have S(vtσ ; Ω) = S(wtσ ; Ω) → S(uσt ; IRn ) = S as R → ∞, uniformly in σ ∈ Σ, 0 ≤ t < 1. In particular, if R ≥ 1 is suﬃciently large, we can achieve that sup S vtσ ; Ω < S1 < 22/n S σ,t

for some constant S1 ∈ IR. Fix such an R and suppose that Ω satisﬁes (1◦ ), (2◦ ) of the Theorem with R2 = 4R = R1−1 for this number R. Let M be the set ∗ |u|2 dx = 1} , M = {u ∈ H01,2 (Ω) ; Ω

and for u ∈ M let

x|∇u|2 dx

F (u) = Ω

denote its center of mass. Suppose (1.1), (1.3) does not admit a positive solution. This is equivalent to the assertion that ∗ 1 1 |∇u|2 dx − ∗ |u|2 dx E(u) = 2 Ω 2 Ω does not admit a critical point u > 0. By Remark 3.2, therefore, E satisﬁes (P.-S.)β on H01,2 (Ω) for n1 S n/2 < β < n2 S n/2 . Equivalently, S0 ( · ; Ω) satisﬁes (P.-S.)β on M for S < β < 22/n S. Moreover, S0 ( · ; Ω) does not admit a critical value in this range. By the deformation lemma Theorem II.3.11 and Remark II.3.12, therefore, for any β in this range there exists ε > 0 and a ﬂow Φ: M × [0, 1] → M such that Φ(Mβ+ε , 1) ⊂ Mβ−ε , where Mβ = {u ∈ M ; S(u; Ω) < β} .

3. The Eﬀect of Topology

193

Given δ > 0, we may cover the interval [S + δ, S1 ] by ﬁnitely many such εintervals and compose the corresponding deformations to obtain a ﬂow Φ: M × [0, 1] → M such that Φ(MS1 , 1) ⊂ MS+δ . Moreover, we may assume that Φ(u, t) = u for all u with S(u; Ω) ≤ S + δ/2. On the other hand, it easily follows either from Theorem I.4.8 or Theorem 3.1 that, given any neighborhood U of Ω, there exists δ > 0 such that F (MS+δ ) ⊂ U . Indeed, for any sequence (um ), where um ∈ MS+ m1 , by Theorem I.4.8 and Theorem 1.2 there exists a subsequence such that, as m → ∞, ∗

|um |2 dx δx(0) , |∇um |2 dx Sδx(0) ¯ Since Ω is smooth we may choose a neighborhood U of for some x(0) ∈ Ω. Ω such that any point p ∈ U has a unique nearest neighbor q = π(p) ∈ Ω and such that the projection π is continuous. Let δ > 0 be determined for such a neighborhood U , and let Φ: M × [0, 1] → M be the corresponding ﬂow constructed above. The map h: Σ × [0, 1] → Ω, given by h(σ, t) = π F Φ(vtσ , 1) , then is well-deﬁned, continuous, and satisﬁes h(σ, 0) = π F Φ(v0 , 1) =: x0 ∈ Ω , h(σ, 1) = σ,

for all σ ∈ Σ

for all σ ∈ Σ .

Hence h is a contraction of Σ in Ω, contradicting (2◦ ). Actually, the eﬀect of topology is much stronger than indicated by Theorem 3.4. In a penetrating analysis, Bahri-Coron [1] have obtained the following result; see also Bahri [2]: 3.5 Theorem. Suppose Ω is a domain in IRn such that Hd (Ω, ZZ2 ) = 0 for some d > 0. Then (1.1), (1.3) admits a positive solution for λ = 0. Note that if Ω ⊂ IR3 is non-contractible then either H1 (Ω, ZZ2 ) or H2 (Ω, ZZ2 )

= 0 and the conclusion of Theorem 3.5 holds. It is conjectured that a similar result will also hold for n ≥ 4. 3.6 Notes. (1◦ ) Benci-Cerami-Passaseo [1] – see, in particular, Theorem 3.1 – have further reﬁned the “method of photography” used in the proof of Theorem 3.4 above and have applied it to obtain multiple solutions of semilinear elliptic equations also on unbounded domains.

194

Chapter III. Limit Cases of the Palais-Smale Condition

(2◦) Schoen-Zhang [1] have obtained pointwise estimates for concentrating solutions um > 0 with um u0 weakly in H 1,2 (S 3 ) to subcritical equations −Δum +

n(n − 2) um = Kupmm −1 on S 3 , 4 n−2

comparing um to a “bubble” Rm2 u∗1 (Rm (x − xm )) of the “right” scale Rm , centered around a suitable point xm . Similar and improved pointwise estimates for bubbling families of solutions um > 0 to ∗

−Δum + hm um = u2m −1 on manifolds have been obtained by Druet-Hebey-Robert [1] in the case when hm → h0 in C 0,α while um u0 weakly in H 1,2 . (3◦) Results similar to Theorem 3.1 have been established for higher order equations of Yamabe type by Hebey-Robert [1]. (4◦) The work of Brezis-Merle [1] Li-Shafrir [1], Adimurthi-Struwe [1] and, ﬁnally, Druet [1] shows that similar geometric quantization patterns also characterize the blow-up behavior for concentrating solutions to semilinear elliptic equations of critical exponential growth on a planar domain. In recent work of Struwe [28] the results of Druet [1] have been extended to critical semilinear elliptic equations of fourth order on four-dimensional domains.

4. The Yamabe Problem Equation (1.1) arises in a geometric context in the problem whether a given metric g0 on a closed manifold M of dimension n ≥ 3 with scalar curvature Rg0 = R0 can be deformed conformally to a metric g of constant scalar curvature. If we let 4 g = u n−2 g0 , where u > 0 gives the conformal factor, the scalar curvature R = Rg of the metric g is given by the equation (4.1)

−

∗ 4(n − 1) Δ0 u + R0 u = Ru2 −1 , n−2

where Δ0 = Δg0 is the Laplace-Beltrami operator on the manifold M with respect to the original metric g0 ; see Yamabe [1], T. Aubin [2], [3]. Observe that by its intrinsic geometric meaning Equation (4.1) is conformally invariant; that is, if u solves (4.1) on (M, g0 ) and if 4

g0 = v n−2 g˜0 ,

v>0

˜ = uv satisﬁes (4.1) on (M, g˜0 ) with is conformal to a metric g˜0 on M , then u ˜ 0 of the metric g˜0 ; see also Aubin [3; R0 replaced by the scalar curvature R Proposition, p. 126].

4. The Yamabe Problem

195

4.1 The sphere. Of particular interest is the case of the sphere M = S n with the standard metric induced by the embedding S n → IRn+1 . We center S n at the origin of IRn × IR. Via the inverse of stereographic projection 2x 1 − |x|2 IRn x → , ∈ S n ⊂ IRn+1 , 1 + |x|2 1 + |x|2 4 S n induces a metric g = [1+|x| 2 ]2 gIRn of constant scalar curvature Rg = n(n−1) n on IR , conformal to the Euclidean metric. The classical solution

u(x) =

[n(n − 2)] [1 + |x|2 ]

n−2 4

n−2 2

of (3.1) on IRn thus reappears in the conformal factor that changes the ﬂat Euclidean metric to a metric of constant scalar curvature, and the invariance of (3.1) under scaling n−2 u → ur,x0 (x) = r 2 u r(x − x0 ) reﬂects the action of the group of conformal diﬀeomorphism of S n via translations and dilations in the chart IRn . The Variational Approach We can give a variational formulation of (4.1) analogous to the one given for (1.1)–(1.3). Let H 1,2 (M, g0 ) be the set of Sobolev functions u: M → IR such that in local coordinates on M , with g0ij (x) denoting the coeﬃcients of the inverse of the metric g0 (x), and with volume element dμ0 = dμg0 = det(g0 ) dx,

we have u2H 1,2 :=

M

|∇u|20 + |u|2 dμ0 < ∞

where |∇u|20 = g0ij ∂i u∂j u. By convention, we tacitly sum over repeated indices. Let cn = 4(n−1) n−2 . In analogy with the Sobolev ratio in the Euclidean case, for 0 = u ∈ H 1,2 (M, g0 ) we deﬁne

c |∇u|20 + R0 |u|2 dμ0 M n . S(u) = S(u; (M, g0)) = 2/2∗ |u|2∗ dμ0 M Then S is of class C 1 on the space H 1,2 (M, g0 ) \ {0}, and critical points u > 0 ∗ of S on the unit sphere in L2 (M, g0 ) correspond to solutions of (4.1) with constant scalar curvature R = S(u), the total scalar curvature of the metric 4 g = u n−2 g0 . In particular, if the Yamabe constant Y0 = Y (M, g0 ) = inf S(u); 0 = u ∈ H 1,2 (M, g0 ) ,

196

Chapter III. Limit Cases of the Palais-Smale Condition

on (M, g0 ) is attained at a function u > 0, Equation (4.1) will hold with R ≡ Y0 . If Y0 = 0, a constant multiple of u then will solve (4.1) with R = 1 or R = −1. Thus, depending on the sign of Y (M, g0 ), the latter question can also be approached as a minimax problem for the free functional ∗ 1 1 2 2 cn |∇u|0 + R0 |u| dμ0 ± ∗ |u|2 dμ0 , E(u) = E(u; (M, g0)) = 2 M 2 M where the “+”-sign (“−”-sign) is valid if Y (M, g0 ) is negative (positive). If Y (M, g0 ) = 0, also R = 0 and (4.1) reduces to a linear equation. Note that the Yamabe constant Y (M, g0 ) is independent of the metric g0 representing a conformal class. In particular, we have n(n − 1) = Y (S n , gS n ) = Y (IRn , gIRn ) = cn S, where S is the Sobolev constant for the ∗ embedding D1,2 (IRn ) → L2 (IRn ). An argument as in Remark I.4.5 then also shows that there holds Y (M, g0 ) ≤ Y (S n , gS n ) for any (M, g0 ). If Y (M, g0 ) < 0, the functional E is coercive and weakly lower semicontinuous on H 1,2 (M, g0 ), as may easily be veriﬁed by the reader. Hence in this case the existence of a non-negative critical point follows from the direct methods; see Chapter I. Alternatively, one may employ the method of sub- and supersolutions to (4.1) used by Kazdan-Warner [1] or Loewner-Nirenberg [1]; see also Sections I.2.3–I.2.6. The diﬃcult case is the case when Y (M, g0 ) > 0. Years after Yamabe’s [1] ﬁrst – unsuccessful – attempt to solve (4.1) for general manifolds (M, g0 ), Trudinger [1; Theorem 2, p. 269] obtained a rigorous existence result for small positive Y (M, g0 ). His approach then was reﬁned by Aubin [2]. By using optimal Sobolev estimates, Aubin was able to show the following result on which Lemma 2.2 above was modeled. 4.2 Lemma. If Y (M, g0 ) < Y (S n , gS n ), then Y (M, g0 ) is attained at a solution u > 0 to (4.1), inducing a conformal metric of constant scalar curvature. As a companion result in the spirit of Lemma 2.3 we ﬁnd that E satisﬁes (P.-S.)β on H 1,2 (M, g0 ) for β < n1 Y (S n , gS n )n/2 . Aubin succeeded in showing that the condition Y (M, g0 ) < Y (S n , gS n ) is satisﬁed if (M, g0 ) is of dimension n ≥ 6 and not locally conformally ﬂat. By means of the “positive mass theorem” Schoen [1], ﬁnally, was able to show that this condition also holds in all the remaining cases when (M, g0 ) = (S n , gS n ). Thus, we obtain the following result. 4.3 Theorem. On any smooth, compact Riemannian manifold (M, g0 ) without boundary of dimension n ≥ 3 there exists a conformal metric of constant scalar curvature, a “Yamabe metric”.

4. The Yamabe Problem

197

The Locally Conformally Flat Case Particularly powerful tools are available for dealing with the Yamabe problem on a locally conformally ﬂat manifold (M, g0 ). In fact, a complete Morse theory for the Yamabe problem may be obtained in this case. 4.4 The developing map. A manifold (M, g0 ) is locally conformally ﬂat if for every point p ∈ M there is a neighborhood U of p in M and a conformal diﬀeomorphism Φ from (U, g0 ) to S n , endowed with the standard metric. Locally conformally ﬂat manifolds were studied by Kuiper and, in particular, by Schoen-Yau [1], whom we follow in the sequel. If n ≥ 3 and if M is simply connected, by a standard monodromy argument the local diﬀeomorphisms Φ: U ⊂ M → S n can be extended to a conformal immersion Φ: M → S n , the developing map, which is unique up to composition with a conformal, or M¨ obius, transformation of S n . Similarly, for an arbitrary locally conformally ﬂat n-manifold (M, g0 ), n ≥ 3, there is a developing map Φ from the universal ˜ is mapped ˜ of M to S n , and the fundamental group π1 (M ) acting on M cover M into the M¨ obius group by a homomorphism called the holonomy representation. One of the main accomplishments in Schoen-Yau [1] is to exhibit a large class of locally conformally ﬂat manifolds where the developing maps are injective. 4.5 The conformal Laplace operator. The left-hand side of (4.1) deﬁnes the conformal Laplace operator on (M, g0 ), L0 = Lg0 = −cn Δ0 + R0 , scaled with the factor cn = 4(n−1) for convenience. Recall that throughout n−2 this section we only consider closed manifolds (M, g0 ), that is, compact manifolds without boundary. Let λ0 be the lowest eigenvalue of L0 and u0 > 0 a 4/(n−2) g0 has a scalar corresponding eigenfunction. By (4.1), the metric g = u0 curvature Rg of constant sign, in fact, of the same sign as λ0 , which is the same as the sign of the Yamabe invariant Y (M, g0 ). Thus, according to the sign of Y (M, g0 ), a closed locally conformally ﬂat manifold (M, g0 ) may be classiﬁed conformally invariantly as scalar positive, scalar negative, or scalar ﬂat. In the following we may restrict ourselves to the case that (M, g0 ) is scalar positive as the most interesting case. For such manifolds we have the following result of Schoen-Yau [1], Proposition 3.3 and 4.4. 4.6 Theorem. Suppose (M, g0 ) is closed, locally conformally ﬂat and scalar ˜ ˜ → S n is injective and Γ = ∂(Φ(M)) positive. Then the developing map Φ: M has vanishing Newtonian capacity. ˜ → M we lift g0 to an equivariant metric g˜0 = π ∗ g0 . Via the covering map π: M Using Theorem 4.6, we can push this metric forward conformally to a complete metric

198

Chapter III. Limit Cases of the Palais-Smale Condition 4

h0 = Φ∗ g˜0 = v0n−2 gS n ˜ ) ⊂ S n , conformal to the standard metric gS n on S n . on Φ(M By a second result of Schoen-Yau [1], Proposition 2.6, the conformal factor ˜ ). v0 diverges uniformly near the boundary Γ of Φ(M ˜ ) there 4.7 Proposition. There exists a constant c > 0 such that for x ∈ Φ(M 2−n holds v0 (x) ≥ c dist(x, Γ ) 2 . The constant c = c(M, g0 ) depends only on the conformal class of g0 and bounds for the curvature of g0 and its derivatives with respect to g0 . As a ﬁnal normalization, upon replacing g0 by a constant multiple of g0 , if necessary, we may assume that (M, g0 ) has unit volume. Similarly, we will impose the volume constraint ∗ (4.2) V ol(g) = dμ = u2 dμ0 = 1. M

M ∗

4

on comparison metrics g = u n−2 g0 , where dμ = dμg = u2 dμ0 . The Yamabe Flow With these prerequisites we now turn to the construction of a Yamabe metric g∞ on (M, g0 ). Instead of using one of the standard variational methods presented so far, following ideas of Hamilton [1] we will approach the problem by means of the parabolic evolution equation corresponding to the “L2 -gradient” ﬂow for the Yamabe energy S(u), whose convergence on a locally conformally ﬂat scalar positive manifold was demonstrated by Ye [1]. 4.8 The evolution problem. The metric g∞ is determined as the limit as t → ∞ of the “Yamabe ﬂow” of metrics (g(t))t≥0 issuing from g(0) = g0 . Letting 4

g(t) = u(·, t) n−2 g0

(4.3)

with Rg(t) = R(t), this ﬂow is deﬁned by letting u: M × [0, ∞[→ IR solve the evolution problem ∂up = sup − L0 u = (s − R)up ∂t

(4.4) for p = 2∗ − 1 =

n+2 n−2

with initial condition u(·, 0) = 1

and with a function s = s(t) determined in such a way that the volume constraint (4.2) is preserved. From the equation

4. The Yamabe Problem

199

p d ∂up V ol(g(t)) = dμ u = u(−L0 u + sup ) dμ0 0 2∗ dt ∂t M M ∗ 2 2 cn |∇u|0 + R0 u dμ0 + s =− u2 dμ0

0=

M

M

we then deduce that (4.5)

s(t) = S(u(t)) for all t.

Multiplying (4.4) by ut and integrating over M , in view of (4.2) we also obtain the energy identity 2s d d 2p up−1 |ut |2 dμ0 + S(u(t)) = ∗ V ol(g(t)) = 0 . dt 2 dt M Hence, the function t → s(t) is non-increasing with a well-deﬁned limit s∞ = lim s(t) ≥ Y (M, g0 ) > 0. t→∞

In addition, we have the a-priori estimate ∞ 2p up−1 |ut |2 dμ0 dt = s(0) − s∞ ≤ S(1) − Y (M, g0 ) < ∞ 0

M

for any smooth global solution u of (4.4), (4.5). In particular, if u together with its space-time derivatives is uniformly bounded on M × [0, ∞ [, a sequence u(t) as t → ∞ will accumulate at a time-independent solution u∞ of (4.4), that is, a solution of (4.1) with R = s∞ . When (M, g0 ) is locally conformally ﬂat the necessary a-priori estimates are a consequence of the following logarithmic gradient bound for equation (4.4), due to Ye [1], Theorem 4. 4.9 Theorem. Suppose that (M, g0 ) is a closed, locally conformally ﬂat nmanifold, n ≥ 3. There exists a constant C = C(M, g0 ) such that for any smooth solution u: M × [0, T [→ IR of (4.4), (4.5) with u(·, 0) = 1 there holds |∇g0 u| ≤C . u M ×[0,T [ sup

By integrating along a shortest geodesic connecting points on M where u(t) achieves its maximum, respectively, its minimum, the estimate of Theorem 4.9 implies the uniform Harnack inequality inf u(t) ≥ c sup u(t) M

M

for 0 < t < T with a uniform constant c > 0. Hence, in view of the volume constraint (4.2), the solution u is uniformly bounded from above and away from 0. But then Equation (4.4) is uniformly parabolic, and we obtain uniform

200

Chapter III. Limit Cases of the Palais-Smale Condition

derivative estimates for u in terms of the data, as well. In particular, any local solution u may be extended for all time and, as t → ∞ suitably, by our remarks above will converge smoothly to a limit u∞ inducing a conformal metric g∞ of constant scalar curvature on M . Theorem 4.9 thus implies the existence of a Yamabe metric on any closed, locally conformally ﬂat manifold (M, g0 ) of dimension n ≥ 3 and completes the proof of Theorem 4.3 in this case. The Proof of Theorem 4.9 (following Ye [1]) We only consider the scalar positive case Y (M, g0 ) > 0. In this case by Theorem 4.6 we can transplant the problem from our abstract manifold M to the sphere via the developing map Φ deﬁned above. We then consider the evolution of the metrics 4 4 ˜(t) n−2 h0 = v(t) n−2 gS n h(t) = Φ∗ g˜(t) = u ˜ ) ⊂ S n , where g˜(t) = π ∗ g(t), u on Φ(M ˜(y, t) = u(π(Φ−1 (y)), t), and where ˜ ), h0 ). v(t) = u ˜(t)v0 . Let Lh0 denote the conformal Laplace operator on (Φ(M ˜ ), h0 ) and (M, g0 ), Since π ◦ Φ−1 deﬁnes a local isometry between (Φ(M from (4.4) we deduce the equation (4.6)

∂u ˜p + Lh 0 u ˜ = s˜ up ∂t

˜ ) × [0, T [. But (4.1) implies on Φ(M ˜ = Rh(t) = v −p LS n v . u ˜−p Lh0 u Hence, in terms of the round spherical metric gS n as background metric on ˜ ) ⊂ S n we can also express (4.4) as the ﬂow equation Φ(M (4.7)

∂v p + LS n v = sv p . ∂t

Moreover, the condition u ˜(·, 0) = u(·, 0) = 1 yields the initial condition (4.8)

v |t=0 = v0 .

˜ ) for the action of the Kleinian group Fix a fundamental domain N ⊂ Φ(M G corresponding to π1 (M ) via the holonomy representation. For the proof of Theorem 4.9 it suﬃces to establish a bound for (4.9)

|∇gS n v| . v N×[0,T [ sup

We may assume that there is an open neighborhood V of the closure of N such ˜) . that dist(V, Γ ) > 0, where Γ = ∂ Φ(M Given q0 ∈ N , we introduce conformal charts for a neighborhood of q0 and for S n \ {q0 }, as follows. After a rotation, we may assume that q0 is the

4. The Yamabe Problem

201

north pole q0 = (0, . . . , 0, 1) ∈ IRn+1 . Then let F : S n \ {q0 } → IRn denote stereographic projection from q0 with inverse 2x |x|2 − 1 −1 F (x) = , , x ∈ IRn . 1 + |x|2 |x|2 + 1 Via stereographic projection from the south pole −q0 , similarly we obtain the coordinate representation 2x 1 − |x|2 , G(x) = 1 + |x|2 1 + |x|2 of S n \ {−q0 }. Note that G(0) = q0 and that we have the relation F −1 (x) = 4 G( |x|x 2 ) for x = 0. We pull back the metric h(t) = v(t) n−2 gS n via F −1 to obtain a conformal metric on IRn , given by (F −1 )∗ h(t) = w(t) n−2 gIRn . 4

Since (F −1 )∗ gS n =

2 1 + |x|2

2 gIRn ,

it follows that w(x, t) =

2 1 + |x|2

n−2 2 x ,t . v G |x|2

Thus, as |x| → ∞ the function w(t) has the asymptotic expansion (4.10) n−2 xi xj 2 2 a i xi n−2 1 a w(x, t) = + + a − δ a + O 0 ij 0 ij |x|n−2 |x|2 2 |x|4 |x|n+1 n−2 1 2 2 a j xj ∂w +O (x, t) = ai − xi (n − 2)a0 + n 2 , ∂xi |x|n |x| |x|n+1 where

∂ v(·, t) ◦ G a0 = a0 (t) = v G(0), t > 0, ai = ai (t) = (0) , ∂xi 1 ∂ 2 v(·, t) ◦ G (0) , 1 ≤ i, j ≤ n . aij = aij (t) = 2 ∂xi ∂xj

The point y(t) with coordinates yi (t) =

ai (t) (n − 2)a0 (t)

is called the center of w(t). Note that if we shift coordinates by y(t), then w(x, ˜ t) = w x + y(t), t has the asymptotic expansion

202

(4.11)

Chapter III. Limit Cases of the Palais-Smale Condition

n−2 2 2 1 w(x, ˜ t) = , a0 + O |x|n−2 |x|n ∂w ˜ (n − 2)2 (x, t) = − ∂xi |x|n

n−2 2

a 0 xi + O

1 |x|n+1

as |x| → ∞. Moreover, a uniform bound for y gives the desired bound for (4.9). Thus, the proof of Theorem 4.9 will be complete once we establish the following estimate. 4.10 Lemma. There is a constant C depending only on (M, g0 ) and dist(V, Γ ), such that |y(t)| ≤ C, uniformly for q ∈ N , 0 ≤ t < T . Proof. Fix some number 0 < t¯ < T . After a rotation of coordinates and a reﬂection in the hyperplane {xn = 0}, if necessary, we may assume that yn (t¯) = maxi |yi (t¯)|. For λ > 0, x = (x , xn ) ∈ IRn denote xλ = (x , 2λ − xn ) the image of x after reﬂection in the hyperplane {xn = λ}. By the expansion (4.10) and the arguments of Gidas-Ni-Nirenberg [1], Lemma 4.2, there exists λ0 ≥ 1 such that for any λ ≥ λ0 there holds (4.12)

w(x, 0) > w(xλ , 0), if xn < λ;

∂w(x, 0) < 0, if xn = λ . ∂xn

Note that λ0 only depends on (M, g0 ) and dist(V, Γ ). Here and in the following we extend v to all of S n × [0, T [ by letting v(x, t) = ∞ for x ∈ Γ , as suggested by Proposition 4.7. Similarly, we extend w to IRn ×[0, T [ by letting w(x, t) = ∞ for x ∈ F (Γ ). In particular, by (4.12) the number λ0 has to be chosen so that F (Γ ) lies “below” the hyperplane {xn = λ0 }; in fact, we may choose λ0 such that F (Γ ) lies strictly below {xn = λ0 }. Since, by assumption, u is smooth on M × [0, ¯ t] and hence v is uniformly smooth on V¯ × [0, ¯ t], the expansions (4.10) are uniform in t ∈ [0, ¯ t]. By the same arguments as cited above, therefore, there exists a number λ1 ≥ λ0 such that for every λ ≥ λ1 and every t ∈ [0, t¯] there holds (4.13)

w(x, t) > w(xλ , t), if xn < λ .

For λ ≥ λ0 deﬁne wλ (x, t) = w(xλ , t). We restrict wλ to the region where / F (Γ ), 0 ≤ t ≤ t¯. Note that the functions w and wλ satisfy xn ≤ λ, x ∈ the evolution equation (4.4) with respect to the ﬂat Euclidean metric on this domain. Moreover, w = wλ for xn = λ. Let I = {λ ∈ IR; λ > λ0 , λ > max yn (t), wλ ≤ w} . 0≤t≤t¯

4. The Yamabe Problem

203

By (4.13) the set I is non-empty. We will show that I is open and (relatively) closed in the interval ]λ0 , ∞ [; hence I =]λ0 , ∞ [, which implies that yn (t¯) ≤ λ0 . Since t¯ < T was arbitrary, this then implies the assertion of the lemma. I is open. Indeed, by (4.12) for λ ≥ λ0 equality wλ ≡ w is impossible. (One might also use the singular set F (Γ ) to rule out wλ ≡ w.) Hence, for any λ ∈ I the parabolic maximum principle and (4.12) imply that we have (4.14)

wλ (x, t) < w(x, t) for xn < λ, 0 ≤ t ≤ t¯ ,

and (4.15)

∂w (x, t) < 0 for xn = λ, 0 ≤ t ≤ t¯ . ∂xn

Moreover, by Proposition 4.7 and uniform boundedness of wλ near F (Γ )×[0, ¯ t], uniformly in λ ≥ λ0 , there exists c > 0 such that there holds lim inf w(x, t) − wλ (x, t) ≥ c > 0 , x→F (Γ )

uniformly in t ∈ [0, t¯] and uniformly in λ ≥ λ0 . By (4.10), given any Λ0 ≥ λ0 there is a constant r0 such that the estimates t] and λ ∈ [λ0 , Λ0 ]. Thus (4.14) and (4.15) hold true for |x| ≥ r0 , for all t ∈ [0, ¯ there exists ε > 0 such that ]λ − ε, λ + ε [⊂ I; that is, I is open. ¯ By continuity, I is also closed. Indeed, suppose that λ > λ0 belongs to I. we have wλ ≤ w and λ ≥ max0≤t≤t¯ yn (t). Suppose that λ = max0≤t≤t¯ yn (t). t]. Shift coordinates by yn (t0 ) and denote by Then λ = yn (t0 ) for some t0 ∈ [0, ¯ ˜ λ = wλ (x , λ + xn , t) = w(x , λ − xn , t), respectively, w(x, ˜ t) = w(x , λ + xn , t), w the transformed functions. Also let Γ˜ be the shifted singular set F (Γ ). Via F we lift w, ˜ w ˜ λ back to the sphere by letting ˜ n−2 gIRn , z n−2 gS n = F ∗ w 4

4

(z λ ) n−2 gS n = F ∗ (w ˜ λ ) n−2 gIRn . 4

4

n n \ F −1 (Γ˜ )) × [0, ¯ t], where S− is the hemisphere Then z, z λ are deﬁned on (S− corresponding to xn < 0. (z is related to v by a conformal diﬀeomorphism of S n ﬁxing the north pole q0 .) Moreover, z and z λ satisfy (4.7). We also know n . that z λ ≤ z and z λ = z along ∂S− ˜ 0) Finally, note that by (4.11) the term an in the expansion (4.10) for w(t vanishes. In terms of z and z λ this translates into the condition

∂z ∂z λ (q0 , t0 ) = (q0 , t0 ) = 0 , ∂ν ∂ν n where ν is the outward unit normal on ∂S− . But then the strong parabolic λ maximum principle implies that z ≡ z , that is, w ≡ wλ . This contradicts (4.12). Hence λ > max0≤t≤t¯ yn (t); that is, λ ∈ I, and I is closed. The proof is complete.

204

Chapter III. Limit Cases of the Palais-Smale Condition

Convergence of the Yamabe Flow in the General Case Ye’s Theorem 4.9 depends heavily on the special symmetries of the sphere. While it is not yet known if a similar result will hold for the Yamabe ﬂow on an arbitrary closed manifold, it is possible to show convergence of the ﬂow in dimensions 3 ≤ n ≤ 5. The following result is due to Schwetlick-Struwe [1] and Brendle [1]. In the case n = 3 a diﬀerent proof was given by Gr¨ uneberg [1]. 4.11 Theorem. Let (M, g0 ) be a smooth compact, scalar positive manifold without boundary of dimension 3 ≤ n ≤ 5. Then there is a unique, global, smooth solution u > 0 to the Yamabe ﬂow (4.4) with initial data u(0) ≡ 1 such 4 that the associated metrics g(t) = u(t) n−2 g0 as t → ∞ converge to a metric g∞ of constant scalar curvature. Schwetlick-Struwe [1] obtain the result for initial metrics g0 with R0 = Rg0 > 0 and whose total scalar curvature s0 is small in the sense that (4.16)

2/n 0 < Y0 ≤ s0 ≤ Y (M, g0 )n/2 + Y (S n , gS n )n/2 .

Brendle [1] later was able to remove this restriction. Moreover, he also treated the case when the initial metric g0 has scalar curvature of varying sign. The proof of Theorem 4.11 for the general case is very technical. In the following we will therefore focus on the case when Rg0 > 0 and when g0 satisﬁes (4.16). Moreover, we may assume that (M, g0 ) is not conformal to (S n , gS n ). To simplify the notation, in the following we rescale time with the factor n+2 so that our ﬂow equation (4.4) becomes p = n−2 (4.17)

ut = (s − R)u , u(0) = 1 .

From (4.1) we also obtain the identity n+2

Rt = u− n−2 L0 ut −

n + 2 ut R n−2 u

for the evolution of the scalar curvature. Thus, using the invariance property n+2

u− n−2 L0 (uv) = Lv = (−cn Δ + R)v , where L = Lg with Δ = Δg , we may rephrase the former equation as (4.18)

Rt = L(s − R) +

n+2 4 R(R − s) = cn ΔR + R(R − s) . n−2 n−2

Finally, in the scaled time variable the energy inequality reads d (4.19) st = R dμ = 2 R(s − R) dμ = −2 |R − s|2 dμ ≤ 0 . dt M M M

4. The Yamabe Problem

205

4.12 Global existence. We readily obtain uniform upper and lower a-priori bounds for the solution u to (4.4) on any ﬁnite time interval. Standard parabolic estimates then yield global existence of the ﬂow. 4s0 ˜ First we show that R > 0 for all t ≥ 0. Let R(t) = e n−2 t R(t). Then, as long as R ≥ 0, from (4.18), (4.19) we have ˜ t − cn ΔR ˜= R

4 ˜ ≥ 0. ˜ ≥ 4R R (R − s + s0 )R n−2 n−2

The maximum principle now implies that ˜ ≥ min R(0) ˜ R = min R0 , M

M

uniformly, and we conclude that 4s0

min R(t) ≥ min R0 · e− n−2 t M

M

as claimed. But then from (4.17) we also obtain the uniform upper bound sup u(t) ≤ es0 t sup u(0) = es0 t M

M

for any t. Moreover, we have 4.13 Lemma. Suppose that u > 0 is a smooth function inducing a metric 4 g = u n−2 g0 with R ≥ 0, normalized by (4.2). Then with constants C, q ≥ 1 depending only on (M, g0 ) there holds C inf u ≥ ||u||1−q L∞ . M

∗

Proof: In view of the equation L0 u = Ru2 −1 ≥ 0, the weak Harnack inequality implies the estimate C inf u ≥ ||u||Lp0 M

with uniform constants C, p0 > 0; see Schwetlick-Struwe [1], Theorem A.2. We may assume p0 ≤ 2∗ ; otherwise, (4.2) gives a uniform lower bound for u and we are done. But then by (4.2) we have ∗

∗

1 = ||u||2L2∗ ≤ ||u||pL0p0 ||u||2L∞−p0 . Hence we obtain

1−2∗ /p0

C inf u ≥ ||u||L∞ M

,

as desired. In view of Lemma 4.13, our uniform upper bound for u(t) then also implies a uniform lower bound for any ﬁnite time. Global existence of the ﬂow (4.4) now follows.

206

Chapter III. Limit Cases of the Palais-Smale Condition

4.14 Curvature decay. For q ≥ 1 also consider the functionals Sq (g) = Rq dμ , Fq (g) = |R − s|q dμ , M

M

so that s(g) = S1 (g). Note that (4.17) and (4.18) give ut q−1 ∗ R Rt dμ + 2 Rq dμ ∂t Sq = q u M M 4q − 2n Rq−1 ΔR dμ + Rq (R − s) dμ = qcn n−2 M M 4(q − 1)cn 4q − 2n q/2 2 =− |∇R | dμ + (Rq − sq )(R − s) dμ . q n−2 M M By convexity, moreover, we have (Rq − sq )(R − s) ≥ |R − s|q+1 . Thus, for 1 ≤ q < n2 we obtain that ∞ ∞ (4.20) Fq+1 (g(t)) dt = 0

0

|R − s|q+1 dμ dt ≤ M

n−2 Sq (g0 ) . 2n − 4q

In particular, for q = 1 we recover the estimate (4.19). Moreover, for q < ﬁnd (4.21)

n 2

we

lim inf Fq+1 (g(t)) = 0 . t→∞

In fact, the convergence is uniform. 4.15 Lemma. For any 1 ≤ p <

n+2 2

there holds Fp (g(t)) → 0 as t → ∞.

The assertion of Lemma 4.15 even holds true for any p < ∞; see SchwetlickStruwe [1], Lemma 3.3. q/p

Proof of Lemma 4.15. By H¨older’s inequality, for 1 ≤ q ≤ p we have Fq ≤ Fp . It therefore suﬃces to prove the claim for p ≥ n+1 2 ≥ 2. For such p we have ut ∂t Fp (g) = p (R − s)p−1 (Rt − st ) dμ + 2∗ |R − s|p dμ u M M 4p p−1 p = pcn (R − s) ΔR dμ + |R − s| R dμ n−2 M M (R − s)p−1 dμ − 2∗ (R − s)p+1 dμ − pst M M 4(p − 1)cn 4p − 2n p/2 2 =− |∇(R − s) | dμ + (R − s)p+1 dμ p n−2 M M 4ps |R − s|p dμ − pst (R − s)p−1 dμ , + n−2 M M

4. The Yamabe Problem

207

where z α := z|z|α−1 for all z ∈ IR, any α > 0. That is, recalling that p ≥ 2 and taking account of (4.19), with constants C = C(p, g0 ) we have cn |∇(R − s)p/2 |2 + R|R − s|p dμ ∂t Fp (g) ≤ − M

(4.22)

+ 2pF2 (g)Fp−1 (g) + ≤ −Y0 Fp∗ (g)

n−2 n

4ps0 |4p − 2n| + s Fp (g) + + 1 Fp+1 (g) n−2 n−2

+ CFp (g) + CFp+1 (g) ,

np where p∗ = n−2 . Here we also used the fact that our assumption Y0 > 0 implies the uniform Sobolev type inequality

n−2 n ∗ cn |∇v|2 + Rv 2 dμ |v|2 dμ ≤

Y0 M

M

for all g = g(t) and all functions v ∈ H (M, g); moreover, we observe that by H¨ older’s inequality we can bound F2 (g)Fp−1 (g) ≤ Fp+1 (g). Furthermore, by H¨ older’s inequality again, there holds 1

1+(1−α)p

Fp+1 (g) = ||R − s||p+1 ≤ ||R − s||pα ||R − s||Lp (M,g) , ∗ Lp+1 (M,g) Lp (M,g) where α =

n 2p

< 1. Applying Young’s inequality

AB ≤ αA1/α + (1 − α)B 1/(1−α) ≤ A1/α + B 1/(1−α) , for any δ > 0 we then obtain the bound p(1+β)

Fp+1 (g) ≤ δ||R − s||pLp∗ (M,g) + C||R − s||Lp (M,g) = δFp∗ (g) α

where C = C(δ) = δ − 1−α , β = we then obtain that

1 . p(1−α)

n−2 n

+ CFp (g)1+β ,

For suﬃciently small δ > 0 from (4.22)

∂t Fp (g) ≤ CFp (g) + CFp (g)1+β and the claim follows from (4.20), (4.21). By means of Lemma 4.15, we can deduce convergence of the ﬂow (4.4) as a consequence of the following proposition. 4.16 Proposition. For any sequence tk → ∞ (k → ∞) there exist constants 0 < γ < 1, C such that for a subsequence there holds 2n (g(tk )) s(tk ) − s∞ ≤ CF n+2

n+2 2n (1+γ)

Indeed, Proposition 4.16 implies the following result.

.

208

Chapter III. Limit Cases of the Palais-Smale Condition

4.17 Lemma. There exist constants 0 < γ < 1, t0 such that for all t ≥ t0 there holds 1+γ s(t) − s∞ ≤ F2 (g(t)) 2 . Proof. We argue indirectly. In view of Lemma 4.15 we thus suppose that for a sequence tk → ∞ (k → ∞) we have s(tk ) − s∞ ≥ F2 (g(tk ))

1+1/k 2

.

But by Proposition 4.16 and H¨ older’s inequality for a subsequence with suitable constants γ > 0, C there holds 2n (g(tk )) s(tk ) − s∞ ≤ CF n+2

n+2 2n (1+γ)

We then conclude that 1 ≤ CF2 (g(tk ))

≤ CF2 (g(tk ))

γ−1/k 2

1+γ 2

.

,

which for large k contradicts Lemma 4.15. The proof is complete. 4.18 Lemma. We have

∞

1

F2 (g(t)) 2 dt < ∞.

0

Proof. In view of (4.19), the assertion of Lemma 4.17 translates into the differential inequality d (s(t) − s∞ ) = −2F2 (g) ≤ −C(s(t) − s∞ )1+δ dt for some 0 < δ < 1. It follows that s(t) − s∞ ≤ Ct−1/δ for some constant C > 0. Hence from (4.19) we deduce 2 2T 2T 1 T F2 (g(t)) 2 dt ≤T F2 (g(t)) dt ≤ (s(T ) − s(2T )) ≤ CT 1−1/δ . 2 T T By diadic decomposition we conclude ∞ ∞ 2k+1 ∞ 1−δ 1 1 F2 (g(t)) 2 dt = F2 (g(t)) 2 dt ≤ C 2− 2δ k < ∞ , 1

k=0

2k

k=0

as claimed. From Lemma 4.18 weak H 1 -convergence of the ﬂow u(t) can be obtained in a straightforward manner. Standard parabolic regularity theory, moreover, yields uniform a-priori bounds for u(t) and its derivatives, which completes the proof of Theorem 4.11; see Brendle [1], Section 3, for details.

4. The Yamabe Problem

209

It thus remains to give the proof of Proposition 4.16. Fix any sequence tk → ∞ (k → ∞) and let uk = u(tk ). From Lemma 4.15 we conclude that the sequence (uk ) is a Palais-Smale sequence for the Yamabe energy, satisfying the conditions R dμ = S(uk ) → s∞ ≥ Y0 sk = M ×{tk }

and

n+2

−cn Δ0 uk + R0 uk = s∞ ukn−2 + o(1) as k → ∞, with error n+2

n+2

o(1) = (Rk − sk )ukn−2 + (sk − s∞ )ukn−2 → 0 in L n+2 → H −1 . 2n

From the general concentration-compactness result Theorem 3.1 for such sequences we then obtain the following result. Let r0 > 0 denote a lower bound for the injectivity radius on (M, g0 ). Fix a function ϕ ∈ C0∞ (Br0 (0)) satisfying ϕ(x) = 1 on Br0 /2 (0) ⊂ IRn and for any x, y ∈ M let ϕy (x) = ϕ(exp−1 y (x)), where exp is the exponential map in the metric g0 . 4.19 Lemma. For any sequence tk → ∞ (k → ∞), letting uk = u(tk ), there exist an integer L and sequences (xk,l )k∈IN , (rk,l )k∈IN , l = 1, . . . , L, such that for a subsequence as k → ∞ there holds uk (x) −

L

2−n

−1 1 ϕxk,l (x) rk,l2 u ¯(rk,l exp−1 xk,l (x)) → u∞ in H (M, g0 ) ,

l=1

where u∞ ≥ 0 solves n+2

n−2 −cn Δ0 u∞ + R0 u∞ = s∞ u∞ on M ,

(4.23)

and where u ¯ is a solution to the equation n+2

¯ = s∞ u ¯ n−2 on IRn . −cn ΔIRn u

(4.24) Moreover, letting

V ol(u∞ ) = M

∗

u2∞ dμ0 , V ol(¯ u) =

∗

u ¯2 dx IRn

for brevity, as k → ∞ we have (4.25)

u). 1 = V ol(uk ) → V ol(u∞ ) + L V ol(¯

By the result of Gidas-Ni-Nirenberg and Obata cited in Remark 3.2, we may assume that n−2 (4n(n − 1)/s∞ ) 4 . u ¯(x) = n−2 (1 + |x|2 ) 2

210

Chapter III. Limit Cases of the Palais-Smale Condition

In particular, u ¯ achieves the Yamabe constant on IRn , which is conformal to n S . Letting

cn IRn |∇v|2gIRn dx ¯ S(v) = n−2 , 2∗ dx n |v| n IR hence we ﬁnd the relation ¯ u) = Y (S n , gS n ) . S(¯ Upon multiplying Equation (4.24) by u ¯ and integrating by parts we then obtain the identity 1−2/2∗ 2/n ¯ u) = s∞ V ol(¯ u) = s∞ V ol(¯ u) . Y (S n , gS n ) = S(¯ Similarly, if u∞ = 0 we may multiply Equation (4.23) by u∞ and integrate by parts to obtain the identity S(u∞ ) = s∞ (V ol(u∞ ))

2/n

.

u), from (4.25) we then obtain Solving for V ol(u∞ ) and V ol(¯ u) = 1 = V ol(u∞ ) + L V ol(¯

(4.26) that is,

S(u∞ ) s∞

n/2 +L

Y (S n , gS n ) s∞

n/2 ,

2/n s∞ = S(u∞ )n/2 + L Y (S n , gS n )n/2 .

But by deﬁnition of the Yamabe invariant we have S(u∞ ) ≥ Y (M, g0 ); moreover, assumption (4.16) and (4.19) imply the estimate 2/n . s∞ ≤ s0 ≤ Y (M, g0 )n/2 + Y (S n , gS n )n/2 Thus, we either have s∞ = s0 and g0 is a Yamabe metric, or L = 0 ; in each case the sequence (uk ) is H 1 -compact with limit u∞ inducing a Yamabe metric. Note that u∞ > 0 by the maximum principle. On the other hand, if u∞ ≡ 0, from (4.16) and (4.19) we obtain 2/n L2/n Y (S n , gS n ) = s∞ < s0 ≤ Y (M, g0 )n/2 + Y (S n , gS n )n/2 . Since Y (M, g0 ) ≤ Y (S n , gS n ) we then conclude that L = 1; that is, at most a single simple bubble develops from (uk ) as k → ∞. Moreover, from (4.26) in this case we have n/2 Y (S n , gS n ) 1= s∞ and it also follows that s∞ = Y (S n , gS n ) = n(n − 1).

4. The Yamabe Problem

211

The Compact Case u∞ > 0 Throughout this part of the argument we follow Brendle [1], Section 6. The operator −

4

ϕ → A∞ ϕ = u∞n−2 L0 ϕ is symmetric with respect to the inner product 4 n−2 u∞ ϕψ dμ0 . (ϕ, ψ)L2 (M,h∞ ) = M

The spectral theorem therefore yields the existence of a complete L2 (M, h∞ )orthonormal sequence of eigenfunctions ϕi , i ∈ IN, of the operator A∞ , with eigenvalues 0 < λi ≤ λi+1 → ∞ as i → ∞. Let i0 be the ﬁrst index such that n+2 and set Z = span{ϕi ; i < i0 }. λi0 > s∞ n−2 Recalling the equation n+2 4 n−2 n−2 A ∞ u∞ − s ∞ u∞ = 0 , = u∞ −cn Δ0 u∞ + R0 u∞ − s∞ u∞

by the implicit function theorem we can ﬁnd a radius ρ > 0 such that for all z = z i ϕi ∈ Z with |z| := ||z||L2 (M,h∞ ) < ρ there exists a unique function uz = u∞ + z + o(|z|) with 4 n−2 (uz − u∞ , ϕi )L2 (M,h∞ ) = u∞ (uz − u∞ )ϕi dμ0 = z i , 1 ≤ i < i0 , M

such that (A∞ uz − s∞ (4.27)

4 uz n−2 uz , ϕi )L2 (M,h∞ ) u∞ n+2 = L0 uz − s∞ uzn−2 ϕi dμ0 = 0

M

for all i ≥ i0 . Moreover, the map z → uz is real analytic. According to results of Lojasiewicz [1], [2], there then exists a number 0 < γ < 1 such that 1+γ ∂ ; (4.28) S(uz ) − S(u∞ ) ≤ C sup i S(uz ) i
212

Chapter III. Limit Cases of the Palais-Smale Condition

Expanding

S(uz ) s∞ − V ol(uz ) V ol(uz ) n−2 n = V ol(uz )

2−2n n

n+2 n−2 L 0 uz − s ∞ uz uz dμ0 ,

M

and recalling (4.27), from (4.28) we can thus estimate 1+γ n+2 n−2 . (4.29) S(uz ) − S(u∞ ) ≤ C sup L 0 uz − s ∞ uz ϕi dμ0 i
M

4.20 Normalization. For any k ∈ IN we choose zk ∈ Z such that (cn |∇(uzk − uk )|20 + R0 |uzk − uk |2 ) dμ0 M = min (cn |∇(uz − uk )|20 + R0 |uz − uk |2 ) dμ0 . |z|<ρ

M

Note that H -convergence uk → u∞ implies that zk → 0 as k → ∞. Decomposing uk = uzk + wk , 1

now

M

∂ L0 wk i uz dμ0 = ∂z

(−cn Δ0 wk + R0 wk ) M

∂ uz dμ0 = 0 ∂z i

for any i < i0 by choice of zk . Thus with error o(1) → 0 as k → ∞ we have 4 n−2 λi u∞ ϕi wk dμ0 = L0 ϕi wk dμ0 M M (4.30) ∂ = L0 wk (ϕi − i uz ) dμ0 = o(1)||wk ||H 1 ∂z M for any such i. 4.21 Lemma. There exist constants c > 0, k0 such that for k ≥ k0 there holds 4 n+2 n−2 u∞ wk2 dμ0 ≤ (1 − c) (cn |∇wk |20 + R0 |wk |2 ) dμ0 . s∞ n−2 M M Proof. Otherwise for a subsequence k → ∞ we may rescale w ˜k = ak wk so that 4 n+2 n−2 2 2 2 s∞ 1= (cn |∇w ˜ k |0 + R0 |w ˜k | ) dμ0 ≤ lim inf u∞ w ˜k dμ0 . k→∞ n−2 M M A subsequence w ˜k → w ˜ = 0 weakly in H 1 , where 4 n+2 n−2 s∞ (cn |∇w| ˜ 20 + R0 |w| ˜ 2 ) dμ0 ≤ u∞ w ˜ 2 dμ0 . n−2 M M

4. The Yamabe Problem

But (4.30) implies that 2 2 (cn |∇w| ˜ 0 + R0 |w| ˜ ) dμ0 ≥ λi0 M

213

4

n−2 u∞ w ˜ 2 dμ0 ,

M

and a contradiction results to the choice of i0 . 4.22 Lemma. There exist constants C > 0, k0 such that for k ≥ k0 there holds ||wk ||H 1 ≤ C

M ×{tk }

|R − s∞ |

2n n+2

n+2 2n dμ .

Proof. By (4.30) there exist functions vk with (vk , ϕi )L2 (M,h∞ ) = 0 for all i < i0 such that ||vk − wk ||H 1 = o(1)||wk ||H 1 , where o(1) → 0 as k → ∞. Recalling the deﬁnition of uz , we then have n+2 (L0 uzk − s∞ uzn−2 )vk dμ0 = 0 . k M

Subtracting the term n+2

n+2

L0 uk − s∞ ukn−2 = (Rg(tk ) − s∞ )ukn−2 in the integrand, from Lemma 4.21 with a uniform constant c > 0 we obtain 4 n+2 n−2 s ∞ u∞ (L0 wk − wk )wk dμ0 c||wk ||2H 1 ≤ n−2 M 4 n+2 n−2 s ∞ u∞ = (L0 wk − wk )vk dμ0 + o(1)||wk ||2H 1 n−2 M n+2 (Rg(tk ) − s∞ )ukn−2 + dk vk dμ0 + o(1)||wk ||2H 1 =: Ik , = M

where n+2

n+2

− dk = s∞ (ukn−2 − uzn−2 k = s∞

4 n + 2 n−2 u∞ wk ) n−2

n+2 n+2 4 4 4 n + 2 n−2 n + 2 n−2 n−2 )wk + s∞ (ukn−2 − uzn−2 − (uzk − u∞ uzk wk ) . k n−2 n−2

Since we assume that n ≤ 5 this term can be bounded 6−n

|dk | ≤ C(uk + |wk | + u∞ ) n−2 (|zk | + |wk |)|wk | . ∗

By Sobolev’s embedding H 1 → L2 , thus we can estimate dk vk dμ0 ≤ o(1)||wk ||2H 1 . M

It follows that

214

Chapter III. Limit Cases of the Palais-Smale Condition

∗

2n

Ik ≤ C ≤C

M ×{tk }

M ×{tk }

|R − s∞ | n+2 u2k dμ0

n+2 2n

||vk ||L2∗ + o(1)||wk ||2H 1

n+2 2n 2n |R − s∞ | n+2 dμ ||wk ||H 1 + o(1)||wk ||2H 1 ,

and hence the claim. Proof of Proposition 4.16 in the case u∞ > 0. Expand n+2

n+2

= L0 uk − Rg(tk ) ukn−2 L0 uzk − s∞ uzn−2 k n+2

n+2

n+2

+ (Rg(tk ) − s∞ )ukn−2 − L0 wk − s∞ (uzn−2 − ukn−2 ) . k Recalling (4.1) and observing the estimates L0 wk ϕi dμ0 = wk L0 ϕi dμ0 M M 4 n−2 = λi u∞ wk ϕi dμ0 ≤ C||wk ||L2∗ ≤ C||wk ||H 1 , 1 ≤ i < i0 , M

together with n+2 n+2 4 n−2 n−2 s∞ (uzk − uk )ϕi dμ0 ≤ C (uk + |wk |) n−2 |wk ||ϕi | dμ0 M

M

≤ C||wk ||L2∗ ≤ C||wk ||H 1 , for any i < i0 we have the bound n+2 n+2 n−2 (L0 uzk − s∞ uzk )ϕi dμ0 ≤ (Rg(tk ) − s∞ )ukn−2 ϕi dμ0 + C||wk ||H 1 . M

M

Lemma 4.22 and (4.29) then yield the estimate (4.31)

S(uzk ) − S(u∞ ) ≤ C

M ×{tk }

|R − s∞ |

2n n+2

n+2 2n (1+γ) dμ .

Likewise we expand L0 uk uk = L0 (uzk + wk )(uzk + wk ) = L0 uzk uzk + 2L0 uk wk − L0 wk wk and use (4.1) to conclude S(uk ) = L0 (uzk + wk )(uzk + wk ) dμ0 M n+2 = L0 uzk uzk dμ0 + 2 Rg(tk ) ukn−2 wk dμ0 − M

M

We may write the latter in the form

M

L0 wk wk dμ0 .

4. The Yamabe Problem

n+2 S(uk ) = s∞ + 2 (Rg(tk ) − s∞ )ukn−2 wk dμ0 M 4 n+2 − (L0 wk wk − wk2 ) dμ0 + Ik , s∞ uzn−2 k n−2 M where

n−2

n−2

Ik = (S(uzk ) − s∞ )V ol(uzk ) n + s∞ (V ol(uzk ) n − 1) n+2 4 n + 2 n−2 uzk wk2 ) dμ0 . (2ukn−2 wk − + s∞ n−2 M

By concavity of the function s → s V ol(uzk )

n−2 n

n−2 n

we have

n−2 n−2 −1≤ (V ol(uzk ) − 1) = n n

M

∗

∗

(u2zk − u2k ) dμ0 .

Therefore we can bound the error term n−2

Ik ≤ (S(uzk ) − s∞ )V ol(uzk ) n n+2 4 n − 2 2∗ n + 2 n−2 n − 2 2∗ uzk − uzk wk2 + 2ukn−2 wk − uk ) dμ0 + s∞ ( n n − 2 n M n−2

= (S(uzk ) − S(u∞ ))V ol(uzk ) n n+2 4 ∗ ∗ n + 2 n−2 2s∞ uzk wk2 /2 − u2zk ) dμ0 . − ∗ (u2k − 2∗ ukn−2 wk + 2∗ 2 n − 2 M Recalling that uk = uzk + wk , we can bound ∗

n+2

|u2k − 2∗ ukn−2 wk + 2∗

4 6−n ∗ n + 2 n−2 uzk wk2 /2 − u2zk | ≤ C(uk + |wk |) n−2 |wk |3 . n−2

In view of (4.31) and Lemma 4.22 we thus obtain the error estimate Ik ≤ C

M ×{tk }

|R − s∞ |

2n n+2

n+2 2n (1+γ) dμ + C||wk ||3H 1

≤C

M ×{tk }

|R − s∞ |

2n n+2

n+2 2n (1+γ) dμ .

By H¨older’s inequality and Lemma 4.22 we also may bound M

n+2 n−2

(Rg(tk ) − s∞ )uk

wk dμ0 ≤

≤C

M ×{tk }

M ×{tk }

|R − s∞ |

2n n+2

|R − s∞ |

2n n+2

n+2 2n dμ ||wk ||L2∗

n+2 n dμ ,

while by Lemma 4.21 with a uniform constant c > 0 for k ≥ k0 we have 4 n+2 s∞ uzn−2 (L0 wk wk − wk2 ) dμ0 ≥ c||wk ||2H 1 . k n−2 M

215

216

Chapter III. Limit Cases of the Palais-Smale Condition

Finally, we can estimate n+2 2n n+2 2n n+2 2n (g(tk )) 2n + s(tk ) − s∞ . |R − s∞ | dμ ≤ F n+2 M ×{tk }

Combining the previous estimates and observing that γ < 1, we obtain 2n (g(tk )) s(tk ) − s∞ = S(uk ) − s∞ ≤ CF n+2

n+2 2n (1+γ)

+ C(s(tk ) − s∞ )1+γ .

Since s(tk ) → s∞ as k → ∞, our claim in Proposition 4.16 follows. The idea of using the Lojasiewicz inequality for proving convergence of a geometric ﬂow was pioneered by Leon Simon [1]. Bubbling: The Case u∞ ≡ 0 . In order to deal with “bubbling” the key idea is to make an optimally balanced choice for the position xk , scale rk , and size αk of the “bubble” separating from the sequence (uk ) by Lemma 4.19, and to use the positive mass theorem. 4.23 The standard bubble. Following Lee-Parker [1], Theorem 5.1, or G¨ unther [2] we introduce local conformal normal coordinates on (M, g0 ). Fix some number N ∈ IN. Given a point x ∈ M , we can ﬁnd a conformal metric 4

hx = ϕxn−2 g0 with ϕx (x) = 1 such that det(D expx (ξ)) = 1 + O(|ξ|N ) near ξ = 0, where D expx (ξ) denotes the diﬀerential of the exponential map in the metric hx at a point ξ ∈ Tx M . Moreover, we have |Rhx (y)| ≤ Cρx (y)2 , where ρx (y) denotes the Riemannian distance from x to y as measured by hx . Also let Gx be the Green’s function of the operator Lhx with pole at x, satisfying Lhx Gx = −cn Δhx Gx + Rhx Gx = 0 for y = x . Moreover, Gx satisﬁes the estimates |Gx (y) − ρx (y)2−n − Ax | ≤ Cρx (y) , |d(Gx (y) − ρx (y)2−n )| ≤ C , with a constant Ax > 0 by the positive mass theorem. As shown by Habermann-Jost [1], Proposition I.1.3, the term Ax smoothly depends on x ∈ M , and we have inf M Ax = A > 0. Fix a smooth cut-oﬀ function χ such that 0 ≤ χ ≤ 1, χ(s) = 1 for s ≤ 1, χ(s) = 0 for s ≥ 2. For any small number δ > 0 scale χδ (s) = χ(s/δ). For ∗ (y), where similar to (2.5) the proﬁle x ∈ M , ε > 0 then let u∗x,ε (y) = ϕx (y)Ux,ε ∗ Ux,ε is given by n−2 4 4n(n − 1)ε2 χδ (ρx (y)) ∗ Ux,ε (y) = + 1 − χ (ρ (y)) G (y) . δ x x n−2 s∞ (ε2 + ρx (y)2 ) 2 Then we have the following result of Schoen [1].

4. The Yamabe Problem

217

4.24 Theorem. Suppose that 3 ≤ n ≤ 5 and that (M, g0 ) is not conformally equivalent to the standard sphere S n . Then, with constants C, c > 0 independent of x ∈ M , δ > 0, and ε > 0 there holds S(u∗x,ε ) ≤ Y (S n , gS n ) − cεn−2 + Cδεn−2 + Cδ −n εn . Proof. A detailed proof, adapting the argument of Schoen [1], thereby using the uniform positivity of the mass, is given by Brendle [1], Proposition B.3. By Theorem 4.24 we can ﬁx δ > 0 such that for suﬃciently small ε > 0 we have S(u∗x,ε ) < Y (S n , gS n ). To be consistent with the notation of Lemma 4.19 henceforth we denote the scale as r > 0 and we let u∗x,r denote a standard bubble of scale r > 0 centered at x. Then Lemma 4.19 provides a decomposition uk = αk u∗xk ,rk + wk =: vk + wk with suitable xk , rk > 0 and αk → 1. In fact, similar to the approach in the case when u∞ > 0 we now normalize this choice by the requirement that (cn |∇(uk − αk u∗xk ,rk )|20 + R0 |uk − αk u∗xk ,rk |2 ) dμ0 M = min (cn |∇(uk − αu∗x,r )|20 + R0 |uk − αu∗x,r |2 ) dμ0 . α>0, x∈M, r>0

M

Then we have the following result similar to (4.30). 4.25 Lemma. With error o(1) → 0 as k → ∞ there holds n+2 vkn−2 ψk wk dμ0 = o(1)||wk ||H 1 M

for the functions ψk = 1 , ψk =

rk exp−1 rk2 xk (x) , ψ = k 2 2 2 rk + d(x, xk ) rk + d(x, xk )2

related to the generators of the family (αu∗x,r )α>0, x∈M, r>0 . Proof. Observe that n+2

n+2

L0 vk = L0 uk − L0 wk = Rg(tk ) ukn−2 − L0 wk = s∞ vkn−2 + dk with n+2

n+2

n+2

dk = (Rg(tk ) − s∞ )ukn−2 + s∞ (ukn−2 − vkn−2 ) − L0 wk → 0 in H −1 . By minimality of the above decomposition with respect to α then we have n+2 0= L0 vk wk dμ0 = s∞ vkn−2 wk dμ0 + o(1)||wk ||H 1 . M

M

as claimed. The other assertions follow similarly.

218

Chapter III. Limit Cases of the Palais-Smale Condition

The following result provides the analogue of Lemma 4.21. It reﬂects the fact that the functions u∗ε deﬁned in (2.5) up to scaling and translation strictly minimize the Sobolev ratio S(u) on IRn . 4.26 Lemma. There exist constants c > 0, k0 such that for k ≥ k0 there holds 4 n+2 vkn−2 wk2 dμ0 ≤ (1 − c) (cn |∇wk |20 + R0 |wk |2 ) dμ0 . s∞ n−2 M M Proof. Otherwise we rescale w ˜k = ak wk so that 4 n+2 (cn |∇w ˜k |20 + R0 |w ˜k |2 ) dμ0 ≤ vkn−2 w ˜k2 dμ0 + o(1) s∞ (4.32) 1 = n−2 M M with error o(1) → 0 for a subsequence k → ∞. Letting n−2

˜k (expxk (rk ξ)): BR/rk (0) ⊂ Txk M → IR , w ˆk (ξ) = rk 2 w for some R < ι0 , the injectivity radius of (M, g0 ), by Sobolev’s inequality we have n−2 n ∗ |w ˆk |2 dξ ≤ 1 + o(1) , cn |∇w ˆk |2 dξ ≤ 1 + o(1) , Y0 BR/rk (0)

BR/rk (0)

1 and a subsequence w ˆk → w ˆ weakly in Hloc (IRn ), where w ˆ ∈ D1,2 (IRn ) satisﬁes w ˆ2 dξ > 0 . 2 2 IRn (1 + |ξ| )

Moreover, passing to the limit in (4.32) and simplifying the factors, we ﬁnd w ˆ2 |∇w| ˆ 2 dξ ≤ n(n + 2) dξ . 2 2 IRn IRn (1 + |ξ| ) But by Lemma 4.25 there holds

1

IRn

ˆ dξ = 0 , n+2 w (1 + |ξ|2 ) 2 1 1 − |ξ|2 w ˆ dξ = 0 , n+2 (1 + |ξ|2 ) 2 1 + |ξ|2

IRn

(1 + |ξ|2 )

IRn

1 n+2 2

ξ w ˆ dξ = 0 . 1 + |ξ|2

A result of Rey [1], Appendix D, pp. 49–51, now implies that w ˆ ≡ 0, which gives the desired contradiction. The proof of Proposition 4.16 now can be completed just as in the case when u∞ > 0.

4. The Yamabe Problem

219

Proof of Proposition 4.16 in the case u∞ ≡ 0. Expand L0 (vk + wk )(vk + wk ) dμ0 S(uk ) = M n+2 = L0 vk vk dμ0 + 2 Rg(tk ) ukn−2 wk dμ0 − L0 wk wk dμ0 M M M n+2 = s∞ + 2 (Rg(tk ) − s∞ )ukn−2 wk dμ0 M 4 n+2 s∞ vkn−2 wk2 ) dμ0 + Ik , − (L0 wk wk − n−2 M where the error term n−2

n−2

Ik = (S(vk ) − s∞ )V ol(vk ) n + s∞ (V ol(vk ) n − 1) n+2 4 n + 2 n−2 vk wk2 ) dμ0 (2ukn−2 wk − + s∞ n−2 M in view of the estimate S(vk ) = S(u∗xk ,rk ) ≤ Y (S n , gS n ) = s∞ may be bounded n+2 4 ∗ ∗ n + 2 n−2 2s∞ vk wk2 /2 − vk2 ) dμ0 ≤ C||wk ||3H 1 (u2k − 2∗ ukn−2 wk + 2∗ Ik ≤ − ∗ 2 n − 2 M similar to the case when u∞ > 0. Since Lemmas 4.26 and 4.19 yield that 4 n+2 (L0 wk wk − s∞ vkn−2 wk2 ) dμ0 ||wk ||2H 1 ≤ C n−2 M for k ≥ k0 , the claim follows from H¨ older’s inequality as before. 4.27 Remarks. (1◦ ) Theorems 4.9 and 4.11 show that it may not always be the best strategy to follow the gradient ﬂow for a variational problem; sometimes an apparently singular (possibly degenerate) evolution equation may be much more tractable, due in this case to the intrinsic geometric meaning of (4.4). We encounter a similar phenomenon in Section 6. (2◦) An argument analogous to Ye’s proof above works in n = 2 dimensions; see Bartz-Struwe-Ye [1]. See also Struwe [26] for a new approach in this case. (3◦) For further material and references on the Yamabe problem we refer the reader to the survey by Lee-Parker [1]. (4◦) See Kazdan-Warner [1], Aubin [3], Bahri-Coron [1], Schoen [2], ChangYang [1], and - more recently - Struwe [27] for the related problem of ﬁnding conformal metrics of prescribed scalar curvature (“Nirenberg’s problem”). (5◦) Another variant of the Yamabe problem is the singular Yamabe problem of ﬁnding complete conformal metrics of constant scalar curvature, for instance, on S n \ Γ , where Γ is a smooth k-dimensional submanifold of S n . If k > n−2 2 , this leads to a coercive problem, reminiscent of the case S(M ) < 0 for the Yamabe problem. This was solved by Loewner-Nirenberg [1]. The case k < n−2 2 is more diﬃcult and was solved only recently by Mazzeo-Smale [1], MazzeoPacard [1].

220

Chapter III. Limit Cases of the Palais-Smale Condition

5. The Dirichlet Problem for the Equation of Constant Mean Curvature Another borderline case of a variational problem is the following: Let Ω be a bounded domain in IR2 with generic point z = (x, y) and let u0 ∈ C 0 (Ω; IR3 ), H ∈ IR be given. Find a solution u ∈ C 2 (Ω; IR3 ) ∩ C 0 (Ω; IR3 ) to the problem Δu = 2Hux ∧ uy u = u0

(5.1) (5.2)

in Ω , on ∂Ω .

Here, for a = (a1 , a2 , a3 ), b = (b1 , b2 , b3 ) ∈ IR3 , a ∧ b denotes the wedge product ∂ u. a ∧ b = (a2 b3 − b2 a3 , a3 b1 − b3 a1 , a1 b2 − b1 a2 ) and, for instance, ux = ∂x (5.1) is the equation satisﬁed by surfaces of mean curvature H in conformal representation. Surprisingly, (5.1) is of variational type. In fact, solutions of (5.1) may arise as “soap bubbles”, that is, surfaces of least area enclosing a given volume. Also for prescribed Dirichlet data, where a geometric interpretation of (5.1) is impossible, we may recognize (5.1) as the Euler-Lagrange equations associated with the variational integral 1 2H 2 |∇u| dz + u · ux ∧ uy dz . EH (u) = 2 Ω 3 Ω For smooth “surfaces” u, the term 1 V (u) := 3

u · ux ∧ uy dz Ω

may be interpreted as the algebraic volume enclosed between the “surface” parametrized by u and a ﬁxed reference surface spanning the “curve” deﬁned by the Dirichlet data u0 ; see Figure 5.1. Indeed, computing the variation of the volume V at a point u ∈ C 2 (Ω; IR3 ) in direction of a vector ϕ ∈ C0∞ (Ω; IR3 ), we obtain 3

d V (u + εϕ)|ε=0 dε ϕ · ux ∧ uy + u · ϕx ∧ uy + u · ux ∧ ϕy dz = Ω =3 ϕ · ux ∧ uy dz + ϕ · (u ∧ uyx + uxy ∧ u) dz , Ω

Ω

and the second integral vanishes by anti-symmetry of the wedge product. Hence critical points u ∈ C 2 (Ω; IR3 ) of E solve (5.1).

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

221

Fig. 5.1. On the volume functional

Small Solutions Since V is cubic, in EH the Dirichlet integral dominates if u is “small” and we can expect that for “small data” and “small” H a solution of (5.1), (5.2) can be obtained by minimizing EH in a suitable convex set. Generalizing earlier results by Heinz [1] and Werner [1], Hildebrandt [2] has obtained the following result, which is conjectured to give the best possible bounds for the type of constraint considered: 5.1 Theorem. that

Suppose u0 ∈ H 1,2 ∩ L∞ (Ω; IR3 ). Then for any H ∈ IR such u0 L∞ · |H| < 1 ,

there exists a solution u ∈ u0 + H01,2 (Ω; IR3 ) of (5.1), (5.2) such that uL∞ ≤ u0 L∞ . The solution u is characterized by the condition EH (u) = min{EH (v) ; v ∈ u0 + H01,2 ∩ L∞ (Ω, IR3 ), vL∞ |H| ≤ 1} . In particular, u is a relative minimizer of EH in u0 + H01,2 ∩ L∞ (Ω; IR3 ). 5.2 Remark. Working with a diﬀerent geometric constraint Wente [1; Theorem 6.1] and Steﬀen [1; Theorem 2.2] prove the existence of a relative minimizer provided 2 E0 (u0 )H 2 < π , 3 where 1 E0 (u) = |∇u|2 dz 2 Ω

222

Chapter III. Limit Cases of the Palais-Smale Condition

is the Dirichlet integral of u. The bound is not optimal, see Struwe [18; Remark IV.4.14]; it is conjectured that E0 (u0 )H 2 < π suﬃces. Proof of Theorem 5.1. Let M = {v ∈ u0 + H01,2 (Ω; IR3 ) ; v ∈ L∞ (Ω; IR3 ), vL∞ |H| ≤ 1} . M is closed and convex, hence weakly closed in H 1,2 (Ω). Note that for u ∈ M we can estimate (5.3)

|H u · ux ∧ uy | ≤

1 1 |H| uL∞ |∇u|2 ≤ |∇u|2 , 2 2

almost everywhere in Ω. Hence EH (u) ≥

1 E0 (u) ; 3

that is, EH is coercive on M with respect to the H 1,2 -norm. Moreover, by (5.3) on M the functional EH may be represented by an integral EH (u) = F (u, ∇u) dx Ω

where F (u, p) =

1 2 2 |p| + Hu · p1 ∧ p2 2 3

is non-negative, continuous in u ∈ IR3 , and convex in p = (p1 , p2 ) ∈ IR3 × IR3 for all u ∈ IR3 with |H||u| ≤ 1. Hence from Theorem 1.6 we infer that EH is weakly lower semi-continuous on M . By Theorem I.1.2, therefore, EH attains its inﬁmum on M at a point u ∈ M . Moreover, EH is analytic in H 1,2 ∩ L∞ (Ω; IR3 ). Therefore we may compute the directional derivative of EH in the direction of any vector pointing from u into M . Let ϕ ∈ C0∞ (Ω), 0 ≤ ϕ ≤ 1 and choose v = u(1 − ϕ) ∈ M as comparison function. Then by (5.3) we have 0 ≥ u − v, DEH (u) = uϕ, DEH (u) ∇u∇(uϕ) + 2H u · ux ∧ uy ϕ dz = Ω 1 |∇u|2 + 2Hu · ux ∧ uy ϕ dz + ∇ |u|2 ∇ϕ dz = 2 Ω Ω 2 1 ≥ ∇ |u| ∇ϕ dz . 2 Ω Hence |u|2 is weakly sub-harmonic on Ω. By the weak maximum principle, Theorem B.6 of Appendix B, |u|2 attains its supremum on ∂Ω. That is, there holds uL∞ ≤ u0 L∞ ,

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

223

as desired. But then u lies interior to M relative to H 1,2 ∩L∞ and DEH (u) = 0; which means that u weakly solves (5.1), (5.2). By a result of Wente [1; Theorem 5.5], ﬁnally, any weak solution of (5.1) is also regular, in fact, analytic in Ω. In view of the cubic character of V , having established the existence of a relatively minimal solution to (5.1), (5.2), we are now led to expect the existence of a second solution for H = 0. This is also supported by geometrical evidence; see Figure 5.2.

Fig. 5.2. A small and a large spherical cap of radius 1/|H| for 0 < |H| < 1 give rise to distinct solutions of (5.1) with boundary data u0 (z) = z on ∂B1 (0; IR2 ).

We start with an analysis of V . The Volume Functional In the preceding theorem we have used the obvious fact that V is smooth on H 1,2 ∩ L∞ (Ω; IR3 ) – but much more is true. Without proof we state the following result due to Wente [1; Section III]: 5.3 Lemma. For any u0 ∈ H 1,2 ∩ L∞ (Ω; IR3 ) the volume functional V extends to an analytic functional on the aﬃne space u0 +H01,2 (Ω; IR3 ) and the following expansion holds: 1 V (u0 + ϕ) = V (u0 ) + ϕ, DV (u0 ) + D2 V (u0 )(ϕ, ϕ) + V (ϕ) . 2 Moreover, the derivatives

(5.4)

∗

DV : H 1,2 (Ω; IR3 ) → H −1 (Ω; IR3 ) = H01,2 (Ω; IR3 ) , 2 ∗ D2 V : H 1,2 (Ω; IR3 ) → H01,2 (Ω; IR3 )

224

Chapter III. Limit Cases of the Palais-Smale Condition

are continuous and bounded in terms of the Dirichlet integral ϕ · ux ∧ uy dz ≤ cE0 (u)E0 (ϕ)1/2 , ϕ, DV (u) = Ω 1/2 u · (ϕx ∧ ψy + ψx ∧ ϕy ) dz ≤ c E0 (u)E0 (ϕ)E0 (ψ) . D2 V (u)(ϕ, ψ) = Ω

Furthermore, DV and D2 V are weakly continuous in the sense that, if um u weakly in H01,2 (Ω; IR3 ), then ϕ, DV (um ) → ϕ, DV (u), for all ϕ ∈ H01,2 (Ω; IR3 ) , D2 V (um )(ϕ, ψ) → D2 V (u)(ϕ, ψ), for all ϕ, ψ ∈ H01,2 (Ω; IR3 ) . Finally, for any u ∈ H 1,2 (Ω; IR3 ) the bilinear form D2 V (u) is compact; that is, if ϕm ϕ, ψm ψ weakly in H01,2 (Ω; IR3 ), then D2 V (u)(ϕm , ψm ) → D2 V (u)(ϕ, ψ) . Wente’s proof is based on the isoperimetric inequality, Theorem 5.4 below. However, with more modern tools, we can also give an entirely analytic proof of Lemma 5.3. In fact, the special properties of V are due to the anti-symmetry of the volume form which gives rise to certain cancellation properties as in our discussion of the compensated compactness scheme; compare Section I.3. In particular, the above bounds for V and its derivatives are related to the observation of Coifman-Lions-Meyer-Semmes [1] that the cross product ϕx ∧ ψy for ϕ , ψ ∈ H01,2 (Ω ; IR3 ) belongs to the Hardy space H1 , with H1 -norm bounded in terms of the Dirichlet integrals of ϕ and ψ. Moreover, we have H 1,2 (Ω; IR3 ) → BM O(Ω; IR3 ), and, by a result of Feﬀerman-Stein [1], BM O is the dual space of H1 . Hence, for instance, we derive the estimate 1/2 u · ϕx ∧ ψy dz ≤ CuBM O ϕx ∧ ψy H1 ≤ C E0 (u)E0 (ϕ)E0 (ψ) . Ω

The remaining properties of V can be derived similarly, using these techniques. The term in our functional EH is simply Dirichlet’s integral E0 (u) =

remaining 1 2 |∇u| dz, well familiar from I.2.7–I.2.10. Both E0 and V are conformally 2 Ω invariant, in particular invariant under scaling u → u(Rx). The fundamental estimate for dealing with the functional E is the isoperimetric inequality for closed surfaces in IR3 ; see for instance Rad´o [2]. This inequality for (5.1), (5.2) plays the same role as the Sobolev inequality Su2L2∗ ≤ u2H 1,2 played for problem (1.1), (1.3). 5.4 Theorem. For any “closed surface” ϕ ∈ H01,2 (Ω; IR3 ) there holds 2 36π V (ϕ) ≤ E0 (ϕ)3 . The constant 36π is best possible.

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

225

Remark. The best constant 36π is attained for instance on the function ϕ ∈ D1,2 (IR2 ), 2 (x, y, 1) , ϕ(x, y) = 1 + x2 + y 2 corresponding to stereographic projection of a sphere of radius 1 above (0, 0) ∈ IR2 onto IR2 , and its rescalings (5.5)

ϕε (x, y) = ϕ(x/ε, y/ε) =

2ε (x, y, ε) . ε2 + x2 + y 2

ϕ and ϕε solve Equation (5.1) on IR2 with H = 1: the mean curvature of the unit sphere in IR3 . Wente’s Uniqueness Result Using the unique continuation property for the analytic Equation (5.1), analogous to Theorem 1.2 we can show that the best constant in the isoperimetric inequality is never achieved on a domain Ω ⊆ IR2 , Ω = IR2 . Moreover, similar to Theorem 1.3, a sharper result holds, due to Wente [4]: 5.5 Theorem. If Ω ⊂ IR2 is smoothly bounded and simply connected then any solution u ∈ H01,2 (Ω; IR3 ) to (5.1) vanishes identically. Proof. By conformal invariance of (5.1) we may assume that Ω is a ball z B1 (0; IR2 ). Reﬂecting u(z) = −u we extend u as a (weak) solution |z|2 2 of (5.1) on IR . From Wente’s regularity result (Wente [1; Theorem 5.5]) we infer that u is smooth and solves (5.1) classically. Now by direct computation we see that the function Φ(x + iy) = |ux |2 − |uy |2 − 2iux · uy is holomorphic on C. Since |∇u|2 dz = 2 IR2

|∇u|2 dz < ∞ ,

B1 (0)

it follows that Φ ∈ L1 (IR ), and hence that Φ ≡ 0 by the mean-value property of holomorphic functions. That is, u is conformal. But then, since u ≡ 0 on ∂B1 (0; IR2 ), it follows that also ∇u ≡ 0 on ∂B1 (0; IR2 ) and hence, by unique continuation, that u ≡ 0; see Hartmann-Wintner [1; Corollary 1]. 2

Theorem 5.5 – like Theorem 1.3 in the context of problem (1.1), (1.3) – proves that EH cannot satisfy (P.-S.) globally on H01,2 (Ω), for any H = 0. Indeed, note that EH (0) = 0. Moreover, by Theorem 5.4 for u ∈ H01,2 (Ω; IR3 ) with 4π E0 (u) = H 2 there holds

226

Chapter III. Limit Cases of the Palais-Smale Condition

;

EH (u) = E0 (u) + 2HV (u) ≥ E0 (u) 1 − ≥ α :=

4H 2 E0 (u) 36π

4π >0, 3H 2

while for any comparison surface u with V (u) = 0, if HV (u) < 0, we have EH (ρu) = ρ2 E0 (u) + 2Hρ3 V (u) → −∞

as ρ → ∞ .

H01,2 (Ω; IR3 ),

Hence, if EH satisﬁed (P.-S.) globally on from Theorem II.6.1 we would obtain a contradiction to Wente’s uniqueness result Theorem 5.5. Local Compactness However, the following analogue of Lemma 2.3 holds: 5.6 Lemma. Suppose u0 ∈ H 1,2 ∩ L∞ (Ω; IR3 ) is a relative minimizer of EH in 4π the space u0 + H 1,2 ∩ L∞ (Ω; IR3 ). Then for any β < EH (u0 ) + 3H 2 condition 1,2 3 (P.-S.)β holds on the aﬃne space {u0 } + H0 (Ω; IR ). For the proof of Lemma 5.6 we need D2 EH (u0 ) to be positive deﬁnite on H01,2 (Ω; IR3 ). By (5.3) this is clear for the relative minimizers constructed in Theorem 5.1, if u0 L∞ |H| < 12 . In the general case some care is needed. Also note the subtle diﬀerence in the topology of H 1,2 ∩ L∞ considered in Theorem 5.1 and H 1,2 considered here. Postponing the proof of Lemma 5.6 for a moment we establish the following result by Brezis and Coron [2; Lemma 3]: 5.7 Lemma. Suppose u0 ∈ H 1,2 ∩ L∞ (Ω; IR3 ) is a relative minimizer of EH in u0 + H01,2 ∩ L∞ (Ω; IR3 ). Then u0 is a relative minimizer of EH in u0 + H01,2 (Ω; IR3 ), and there exists a constant δ > 0 such that D2 EH (u0 )(ϕ, ϕ) ≥ δ E0 (ϕ), for all ϕ ∈ H01,2 (Ω; IR3 ) . Proof. We may assume that H = 0. By density of C0∞ (Ω; IR3 ) in H01,2 (Ω; IR3 ), since u0 is a relative minimizer in u0 + H01,2 ∩ L∞ (Ω; IR3 ) clearly we have δ = inf D2 EH (u0 )(ϕ, ϕ) ; ϕ ∈ H01,2 (Ω; IR3 ), E0 (ϕ) = 1 ≥ 0 . Note that D2 EH (u0 )(ϕ, ϕ) = 2E0 (ϕ) + 2HD2 V (u0 )(ϕ, ϕ) , and D2 V (u0 ) is compact by Lemma 5.3. Hence, if ν := 2H

inf

ϕ∈H01,2 (Ω;IR3 ); E0 (ϕ)=1

D2 V (u0 )(ϕ, ϕ) < 0 ,

then ν is attained. Indeed, a minimizing sequence (ϕm ) for ν accumulates weakly in H01,2 (Ω; IR3 ) at a limit ϕ with

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

227

2HD2 V (u0 )(ϕ, ϕ) = ν < 0 , E0 (ϕ) ≤ 1 . In particular, ϕ = 0, and we can consider the normalized function ϕ¯ = √

ϕ E0 (ϕ)

as comparison function with ¯ ϕ) ¯ = 2HD2 V (u0 )(ϕ,

ν . E0 (ϕ)

It follows that E0 (ϕ) = 1; that is, ν is attained. In terms of δ, we thus conclude that, if δ = 2 + ν < 2, then δ is attained. In particular, if δ = 0 there exists ϕ such that E0 (ϕ) = 1 and D2 EH (u0 )(ϕ, ϕ) = 0 = inf D2 EH (u0 )(ψ, ψ) ; ψ ∈ H01,2 (Ω; IR3 ), E0 (ψ) = 1 . Necessarily, ϕ satisﬁes the Euler equation for D2 EH (u0 ): Δϕ = 2H(u0x ∧ ϕy + ϕx ∧ u0y ) . It follows from a result of Wente [5; Lemma 3.1 ] that ϕ ∈ L∞ (Ω; IR3 ); see also Brezis-Coron [2; Lemma A.1] or Struwe [18; Theorem III.5.1]. Hence by minimality of u0 in u0 + H01,2 ∩ L∞ (Ω; IR3 ), for small |t| from Lemma 5.3 we have EH (u0 ) ≤ EH (u0 + tϕ) = EH (u0 ) + tϕ, DEH (u0 ) t2 2 D EH (u0 )(ϕ, ϕ) + 2Ht3 V (ϕ) 2 = EH (u0 ) + 2Ht3 V (ϕ) , +

and it follows that V (ϕ) = 0; that is, EH (u0 + tϕ) = EH (u0 ) for all t ∈ IR. But then for small |t| also u0 + tϕ is a relative minimizer of EH and satisﬁes (5.1), (5.2): Δ(u0 + tϕ) = 2H(u0 + tϕ)x ∧ (u0 + tϕ)y . Since H = 0, diﬀerentiating twice with respect to t gives

But then D2 V (u0 )(ϕ, ϕ) = 2

Ω

0 = ϕx ∧ ϕy . u0 · ϕx ∧ ϕy dz = 0, and we obtain

δ = d2 EH (u0 )(ϕ, ϕ) = 2E0 (ϕ) = 2 , contrary to the assumption about δ. Proof of Lemma 5.6. Let (um ) be a (P.-S.)β -sequence in u0 + H01,2 (Ω; IR3 ). Consider ϕm = um − u0 and note that by Lemma 5.3 EH (um ) = EH (u0 + ϕm ) = EH (u0 ) + ϕm , DEH (u0 ) 1 + D2 EH (u0 )(ϕm , ϕm ) + 2HV (ϕm ) 2 1 = EH (u0 ) + D2 EH (u0 )(ϕm , ϕm ) + 2HV (ϕm ) , 2 while

228

Chapter III. Limit Cases of the Palais-Smale Condition

1 o(1) E0 (ϕm ) 2 = ϕm , DEH (um ) = ϕm , DEH (u0 ) + D2 EH (u0 )(ϕm , ϕm ) + 6HV (ϕm ) = D2 EH (u0 )(ϕm , ϕm ) + 6HV (ϕm ) , where o(1) → 0 (m → ∞). Subtracting, we obtain 1 D2 EH (u0 )(ϕm , ϕm ) = 6 EH (um ) − EH (u0 ) + o(1) E0 (ϕm ) 2 , and by Lemma 5.7 it follows that (ϕm ) and hence (um ) is bounded. We may assume that um u weakly. Then for any ϕ ∈ H01,2 (Ω; IR3 ) by Lemma 5.3 also o(1) = ϕ, DEH (um ) → ϕ, DEH (u) , and it follows that u solves (5.1), (5.2). Thus, t = 1 is a critical point for the cubic function t → EH (u0 + t(u − u0 )) which also attains a relative minimum at t = 0, and we conclude that EH (u) ≥ EH (u0 ) . Now let ψm = um − u and note that by Lemma 5.3 we have EH (um ) = EH (u) + E0 (ψm ) + HD2 V (u)(ψm , ψm ) + 2HV (ψm ) = EH (u) + EH (ψm ) + HD2 V (u)(ψm , ψm ) = EH (u) + EH (ψm ) + o(1) , o(1) = ψm , DEH (um ) = ψm , DEH (ψm ) + 2HD2 V (u)(ψm , ψm ) = ψm , DEH (ψm ) + o(1) , with error o(1) → 0 as m → ∞. That is, for m suﬃciently large EH (ψm ) = E0 (ψm ) + 2HV (ψm ) ≤ EH (um ) − EH (u) + o(1) 4π ≤ EH (um ) − EH (u0 ) + o(1) ≤ c < , 3H 2 while

1 ψm , DEH (ψm ) = E0 (ψm ) + 3HV (ψm ) = o(1) . 2 It follows that for m ≥ m0 there holds 4π E0 (ψm ) ≤ c < 2 , H whence by the isoperimetric inequality Theorem 5.4 we have o(1) = E0 (ψm ) + 3HV (ψm ) ; H 2 E0 (ψm ) ≥ cE0 (ψm ) ≥ E0 (ψm ) 1 − 4π for some c > 0, and ψm → 0 strongly, as desired.

5. The Dirichlet Problem for the Equation of Constant Mean Curvature

229

Large Solutions We can now state the main result of this section, the existence of “large” solutions to the Dirichlet problem (5.1), (5.2), due to Brezis-Coron [2] and Struwe [12], [11], with a contribution by Steﬀen [2]. 5.8 Theorem. Suppose u0 ≡ const., H = 0 and assume that EH admits a relative minimum u on u0 + H01,2 (Ω; IR3 ). Then there exists a further solution u, distinct from u, of (5.1), (5.2). Proof. We may assume that u = u0 . Then by Lemma 5.6 the theorem follows from Theorem II.6.1 once we establish the following two conditions: (1◦) There exists u1 such that EH (u1 ) < EH (u0 ). (2◦) Letting P = p ∈ C 0 [0, 1]; u0 + H01,2 (Ω; IR3 ) ; p(0) = u0 , p(1) = u1 , we have β = inf sup EH (u) < EH (u0 ) + p∈P u∈p

4π . 3H 2

Conditions (1◦ ) and (2◦ ) will be established by using the sphere-attaching mechanism of Wente [2; p. 285 f.], [3], reﬁned by Brezis and Coron [2; Lemma 5]: Since u = u0 ≡ const., there exists some point z0 = (x0 , y0 ) ∈ Ω such that ∇u(z0 ) = 0. By conformal invariance of (5.1) we may assume that z0 = 0, u(z0 ) = 0, and that with ux (0) = a = (a1 , a2 , a3 ), uy = b = (b1 , b2 , b3 ) there holds H(a1 + b2 ) < 0 . For ε > 0 now let ϕε (x, y) = ε2 +x2ε2 +y2 (x, y, ε) be the stereographic projection of IR2 onto a sphere of radius 1 centered at (0, 0, 1) considered earlier, and let ξ ∈ C0∞ (Ω) be a symmetric cut-oﬀ function such that ξ(z) = ξ(−z) and ξ ≡ 1 in a neighborhood of z0 = 0. Deﬁne ut = uεt := u + tξϕε ∈ u + H01,2 (Ω; IR3 ) . For small ε > 0 the surface ut “looks” like a sphere of radius t attached to u above u(0). Now compute, using (5.4), t2 2 D EH (u)(ξϕε , ξϕε ) + 2Ht3 V (ξϕε ) 2 = EH (u) + t2 E0 (ξϕε ) + 2Ht3 V (ξϕε ) u · (ξϕε )x ∧ (ξϕε )y dx dy . + 2Ht2

EH (ut ) = EH (u) +

Ω

230

Chapter III. Limit Cases of the Palais-Smale Condition

Clearly E0 (ξϕε ) ≤ E0 (ϕε ; IR2 ) + O(ε2 ) = 4π + O(ε2 ) , while V (ξϕε ) = V (ϕε ; IR2 ) + O(ε3 ) =

4π + O(ε3 ) . 3

Expand u(x, y) = u(0) + ux (0)x + uy (0)y + O(r 2 ) = ax + by + O(r 2 ) , where r 2 = x2 + y 2 . Upon integrating by parts and using anti-symmetry of the wedge product, we obtain the following expression: (ax + by) · (ξϕε )x ∧ (ξϕε )y dx dy 2H Ω a ∧ (ξϕε )y + (ξϕε )x ∧ b · (ξϕε ) dx dy =H Ω 4ε2 a ∧ (0, 1, 0) + (1, 0, 0) ∧ b · ξ 2 2 (x, y, ε) dx dy =H (ε + r 2 )2 Ω ε3 = 4H(a1 + b2 ) ξ 2 dx dy 2 2 2 Ω (ε + r ) 4ε2 (a3 x + b3 y) 2 ξ 2 dx dy . − 4H (ε + r 2 )2 Ω Since ξ is symmetric the last term vanishes, while for suﬃciently small ε > 0 ε3 1 επ 2 ξ dx dy ≥ dx dy = >0. 2 2 2 4ε Bε (0) 4 Ω (ε + r ) Finally, since |∇ϕε (x, y)| ≤

cε ε2 +x2 +y 2 ,

we may estimate

O(r 2 ) · (ξϕε )x ∧ (ξϕε )y dx dy Ω O(r 2 )|∇ϕε |2 dx dy + O(ε2 ) ≤ Ω ε2 r 2 dx dy + O(ε2 ) ≤c 2 2 2 Ω (ε + r ) ε2 ≤c dx dy + c dx dy + O(ε2 ) 2 Bε (0) Ω\Bε (0) r ≤ cε2 + cε2 |ln ε| . Hence, for t ≥ 0 we may estimate 4π EH (ut ) ≤ EH (u) +t2 4π +H(a1 +b2 )πε+cε2 |ln ε| +cε2 +2Ht3 +O(ε3 ) . 3

6. Harmonic Maps of Riemannian Surfaces

231

In particular, if H < 0 and if ε > 0 is chosen suﬃciently small, there exists T > 0 such that uT satisﬁes EH (uT ) < EH (u) and, moreover, sup EH (ut ) < EH (u) + 0≤t≤T

4π , 3H 2

as desired. The case H > 0 follows in a similar way by considering (ut )t≤0 . 5.9 Remarks. A “global” compactness condition analogous to Theorem 3.1 also holds in the case of Equation (5.1); see Brezis-Coron [3], Struwe [11]. As a consequence, it is not hard to see the analogue of Theorem 3.4 (Rentzmann [1], Patnaik [1]; if Ω is an annulus the existence of non-trivial solutions u ∈ H01,2 (Ω; IR3 ) to (5.1) was already shown by Wente [4]). It would be interesting to further investigate the eﬀect of topology, for instance, in the spirit of BahriCoron [1]. For further material and references on the Dirichlet and Plateau problems for surfaces of prescribed mean curvature we refer the interested reader to the lecture notes by Struwe [18]. A result analogous to Theorem 5.8 for variable mean curvature H = H(u) has recently been obtained by Struwe [19]; see also Wang, G.F. [1].

6. Harmonic Maps of Riemannian Surfaces As our ﬁnal example we now present a borderline variational problem in a nonsmooth setting, where critical points under suitable conditions can be obtained via a ﬂow approach similar to Section 4. Given a smooth, closed Riemannian surface Σ with metric γ and any smooth, closed k-manifold N with metric g, a natural generalization of Dirichlet’s integral for C 1 -functions on a domain in IRn is the energy functional e(u) dμ . E(u) = Σ

Here, in local coordinates on Σ and N , the energy density e(u) is given by e(u) =

1≤α,β≤2 1≤i,j≤k

1 αβ ∂ i ∂ j γ (x)gij (u) u u , 2 ∂xa ∂xβ

with γ αβ = (γαβ )−1 denoting the coeﬃcients of the inverse of the matrix (γαβ ) representing the metric γ, (gij ) representing g, and with area element dμ = |γ| dx, |γ| = det(γαβ ) . Since we assume that both Σ and N are compact, this expression may be simpliﬁed considerably: First, by the Nash embedding theorem, see for instance

232

Chapter III. Limit Cases of the Palais-Smale Condition

Nash [1] or Schwartz [2; pp. 43–53], any compact manifold N may be isometrically embedded into some Euclidean IRn . A new proof – avoiding the “hard” implicit function theorem – has recently been obtained by G¨ unther [1]. Moreover, E is invariant under conformal mappings of Σ. Thus, by the uniformization theorem, we may assume that either Σ = S 2 or Σ = T 2 = IR2 /ZZ2 , or is a quotient of the upper half-space IH, endowed with the hyperbolic metric y1 dx dy. In particular, if Σ = T 2 , the energy density is simply given by e(u) = 12 |∇u|2 and E becomes the standard Dirichlet integral for mappings u: T 2 = IR2 /ZZ2 → N ⊂ IRn . Consider the space C 1 (Σ; N ) of C 1 -functions u: Σ → N → IRn . Note that C 1 (Σ; N ) is a manifold with tangent space given by Tu C 1 (Σ; N ) = ϕ ∈ C 1 (Σ; IRn ) ; ϕ(x) ∈ Tu(x) N for x ∈ Σ , and E is diﬀerentiable on this space. In fact, if we consider variations ϕ ∈ Tu C 1 (Σ; N ) such that supp(ϕ) is contained in a chart U on Σ whose image u(U ) is contained in a coordinate patch of N , then in order to compute the variation of E in direction ϕ it suﬃces to work in one coordinate frame – both on Σ and N – and all computations can be done as in the “ﬂat” case. From this, the diﬀerentiability of E is immediate and the following deﬁnition is meaningful. 6.1 Definition. map.

A stationary point u ∈ C 1 (Σ; N ) of E is called a harmonic

The concept of harmonic map generalizes the notion of (closed) geodesic to higher dimensions; compare Section II.4. Moreover, if we choose N = IRn , we see that harmonic functions simply appear as special cases of harmonic maps. However, when the target space is curved, the Euler-Lagrange equations for harmonic maps no longer will be linear, as we shall presently see. The Euler-Lagrange Equations for Harmonic Maps First consider the model case Σ = T 2 = IR2 /ZZ2 , where E reduces to the standard Dirichlet integral for doubly periodic mappings u: IR2 → N ⊂ IRn , restricted to a fundamental domain. In this case, if u: T 2 = IR2 /ZZ2 → N ⊂ IRn is harmonic of class C 2 , the ﬁrst variation of E gives 0 = ϕ, DE(u) = ∇u∇ϕ dx = − Δu ϕ dx , T2

T2

for all doubly periodic ϕ ∈ C 1 (IR2 ; IRn ) satisfying the condition ϕ(x) ∈ Tu(x) N for all x ∈ IR2 .

6. Harmonic Maps of Riemannian Surfaces

233

That is, −Δu(x) is orthogonal to the tangent space of N at the point u(x), for any x ∈ T 2 ; in symbols: −Δu(x) =

n

λi (x)νi (u(x)) ⊥ Tu(x) N

for all x ∈ T 2 ,

i=k+1

where νk+1 , . . . , νn is a smooth local orthonormal frame for the normal bundle T N ⊥ near u(x), and where λk+1 , . . . , λn are scalar functions. In general, the Laplace operator will be replaced by the Laplace-Beltrami operator ΔΣ on Σ. The particular structure of this equation can best be seen when N = S n−1 ⊂ IRn . In this case, if u: T 2 → S n−1 ⊂ IRn is of class C 2 and harmonic, it follows that −Δu = λu for some continuous function λ: T 2 → IR. Testing this relation with u and noting that |u| ≡ 1, u · ∇u ≡ 0, we see that λ = −Δu · u = − div(u · ∇u) + |∇u|2 = |∇u|2 ; that is, harmonic maps into spheres satisfy the relation −Δu = u|∇u|2 . For general target manifolds N we may proceed similarly. Fix an index i ∈ {k + 1, . . . , n}. Then, since ∂α u = ∂x∂α u ∈ Tu N for any α, we have λi = −Δu, νi ◦ u = − div∇u, νi ◦ u + ∇u, (dνi ◦ u) · ∇u = ∇u, (dνi ◦ u) · ∇u , where ·, · denotes the scalar product. That is, we have −Δu = A(u)(∇u, ∇u) , where A(u): Tu N × Tu N → Tu⊥ N denotes the second fundamental form of N , given by A(p)(ξ, η) =

n

νi (p)Ai (p)(ξ, η) ,

Ai (p)(ξ, η) = ξ, dνi(p)η ,

i=k+1

for p ∈ N , ξ, η ∈ Tp N . For general domains, similarly the harmonic map equation reads −ΔΣ u = A(u)(∇u, ∇u)Σ ⊥ Tu N , with

A(u)(∇u, ∇u)Σ =

γ αβ A(u)(∂α u, ∂β u) .

α,β

Note that by compactness of N the coeﬃcients of the form A are uniformly bounded.

234

Chapter III. Limit Cases of the Palais-Smale Condition

Bochner Identity Upon diﬀerentiating the harmonic map equation and taking the scalar product with the components of ∇u, we obtain the equation 1 −Δe(u) + |∇2 u|2 = − Δ |∇u|2 + |∇2 u|2 2 = −∇(Δu) · ∇u = ∇ A(u)(∇u, ∇u) · ∇u = div A(u)(∇u, ∇u) · ∇u − A(u)(∇u, ∇u) · Δu = |A(u)(∇u, ∇u)|2 ≤ Ce(u)2 for the energy density e(u). In the model case of harmonic maps u: T 2 = IR2 /ZZ2 → S n−1 → IRn , this identity simply takes the form −Δe(u) + |∇2 u|2 = |∇u|4 = 4e(u)2 . For a general domain manifold Σ, additional terms related to the curvature of Σ appear. If we do not care about the precise form of these terms we then obtain the inequality −ΔΣ e(u) + |∇du|2Σ ≤ C(1 + e(u))e(u) with a constant C depending on N and Σ. The most interesting variant of the Bochner identity results if we work intrinsically on the manifold N and use covariant diﬀerentiation instead of taking the ordinary gradient. Then we obtain the diﬀerential inequality −ΔΣ e(u) ≤ κN e(u)2 + Ce(u) for the energy density of u, where κN ≥ 0 denotes an upper bound for the sectional curvature of N , and where C denotes a constant depending only on Σ and N . See for instance Jost [3], formula (3.2.10). With these remarks we hope that the reader has become somewhat familiar with the concept of harmonic maps. Moreover, to ﬁx ideas, in the following one may always think of mappings u: T 2 = IR2 /ZZ2 → S n−1 ⊂ IRn . In this special case already, all essential diﬃculties appear and nothing of the ﬂavour of the results will be lost. The Homotopy Problem and its Functional Analytic Setting A natural generalization of Dirichlet’s problem for harmonic functions is the following “Homotopy Problem”: Given a map u0 : Σ → N , is there a harmonic map u homotopic to u0 ? As in the case of a scalar function u: Ω → IR we may attempt to approach this problem by direct methods. Denote

6. Harmonic Maps of Riemannian Surfaces

235

H 1,2 (Σ; N ) = u ∈ H 1,2 (Σ; IRn ) ; u(x) ∈ N for a.e. x ∈ Σ the space of H 1,2 -mappings into N . (If Σ = T 2 , then H 1,2 (Σ; N ) is the space 1,2 (IR2 ; N ) of period 1 in both variables, restricted to a funof mappings u ∈ Hloc damental domain.) Note that E is weakly lower semi-continuous and coercive on H 1,2 (Σ; N ) with respect to H 1,2 (Σ; IRn ). Moreover, by a result of Schoen-Uhlenbeck [2; Section IV] we have: 6.2 Theorem. (1◦ ) The space C ∞ (Σ; N ) of smooth maps u: Σ → N ⊂ IRn is dense in H 1,2 (Σ; N ). (2◦) The homotopy class of any map u0 ∈ H 1,2 (Σ; N ) is well-deﬁned. Proof. (1◦ ) The argument basically following

is due to Courant [1; p. 214 f.]. Let ϕ ∈ C0∞ B1 (0) satisfy 0 ≤ ϕ ≤ 1, B1 (0) ϕ dx = 1, and for ε > 0 let ϕε = ε−2 ϕ xε . Given u ∈ H 1,2 (Σ; IRn ), denote u ∗ ϕε (x0 ) = u(x)ϕε (x0 − x) dx Bε (x0 )

its molliﬁcation with ϕε (in local coordinates on Σ). Note

that u ∗ ϕε ∈ C ∞ (Σ; IRn ) and u ∗ ϕε → u in H 1,2 (Σ; IRn ) as ε → 0. Let – denote average and let dist(P, N ) = inf{|P − Q| ; Q ∈ N } denote the distance of a point P from N . For x0 ∈ Σ, ε > 0 and y ∈ Bε (x0 ) now estimate 2 u(x) − u(y) ϕε (x0 − x) dx dist2 u ∗ ϕε (x0 ), N ≤ Bε (x0 ) u(x) − u(y)2 dx . ≤C— Bε (x0 )

Taking the average with respect to y ∈ Bε (x0 ), therefore we can bound — |u(x) − u(y)|2 dx dy dist2 u ∗ ϕε (x0 ), N ≤ C — Bε (x0 )

Bε (x0 )

1 ∇u y + t(x − y) 2 dt dx dy — ≤ Cε2 — Bε (x0 ) Bε (x0 ) 0 ≤C |∇u|2 dx . Bε (x0 )

By absolute continuity of the Lebesgue integral, the latter is small for small ε, uniformly in x0 . In particular, for ε < ε0 = ε0 (u, N ) the distance of uε = u ∗ ϕε from N is smaller than d/2, where d = d(N ) > 0 is the focal distance from N . We then can smoothly project uε down to N to obtain a function u ˜ε ∈ C ∞ (Σ; N ), satisfying ˜ε ) ⊥ Tu˜ε N , uε − u ˜ε L∞ → 0 as ε → 0 . (uε − u

236

Chapter III. Limit Cases of the Palais-Smale Condition

Since uε → u in H 1,2 as ε → 0, uε − uε L2 + uε − uL2 → 0 as ε → 0 . ˜ uε − uL2 ≤ ˜ Moreover, representing uε locally as uε = u ˜ε + i λiε νi ◦ u ˜ε , by orthogonality we have uε | 2 . uε + λiε ∇(νi ◦ u ˜ε )|2 ≥ 1 − C λiε L∞ |∇˜ |∇uε |2 ≥ |∇˜ i

i

Hence lim sup ∇˜ uε L2 ≤ lim ∇uε L2 = ∇uL2 , ε→0

ε→0

and it follows that u ˜ε → u in H (Σ; N ), as desired. (2◦) Deﬁne the homotopy class of u ∈ H 1,2 (Σ; N ) as the homotopy class of any map v ∈ C ∞ (Σ; N ) with v − uH 1,2 ≤ δ for some δ = δ(u, N ) > 0 to be determined. To see that this is well-deﬁned, let v0 , v1 ∈ C ∞ (Σ; N ) satisfy vi − uH 1,2 ≤ δ, i = 0, 1, and let vt = (1 − t)v0 + tv1 be a homotopy connecting v0 and v1 in C ∞ (Σ; IRn ). Note that 1,2

vt − uH 1,2 ≤ (1 − t)v0 − uH 1,2 + tv1 − uH 1,2 ≤ δ for 0 ≤ t ≤ 1. Moreover, for any w ∈ H 1,2 , ε > 0 we can estimate w ∗ ϕε L∞ ≤ Cw ∗ ϕε H 2,2 ≤ C(1 + ε−1 )wH 1,2 . Thus for 0 < ε0 = ε0 (u, N ) ≤ 1 as determined in (1◦ ) and 0 ≤ t ≤ 1 we have dist(vt ∗ ϕε0 , N ) ≤ dist(u ∗ ϕε0 , N ) + (vt − u) ∗ ϕε0 L∞ ≤ dist(u ∗ ϕε0 , N ) + C1 ε−1 0 vt − uH 1,2 ≤ d/2 + C1 δ/ε0 ≤ 3d/4 , provided that we choose 4C1 δ ≤ dε0 . We may then project the maps vt ∗ ϕε0 to maps vt : ∗ ϕε0 ∈ C ∞ (Σ; N ), 0 ≤ t ≤ 1, connecting v0 : ∗ ϕε0 and v1 : ∗ ϕε0 . For i = 0, 1 and 0 < ε < ε0 (u, N ), ﬁnally, as in (1◦ ) we can estimate 2 2 |∇vi | dx ≤ C |∇u|2 dx + Cδ ≤ 3d/4 , dist (vi ∗ ϕε (x0 ), N ) ≤ C Bε (x0 )

Bε (x0 )

provided δ = δ(u, N ) > 0 is chosen suﬃciently small. Thus v0 and v0 : ∗ ϕε0 are ∗ ϕε , 0 ≤ ε ≤ ε0 , and similarly for v1 , showing that v0 homotopic through v0: and v1 belong to the same homotopy class. Bethuel [1] showed that on an m-ball B for any m > 2 the space C ∞ (B; N ) is dense in H 1,2 (B; N ) if and only if π2 (N ) = 0. For general domain manifolds, Bethuel’s work was completed by Hang-Lin [1], after they uncovered a mistake in Bethuel’s original work dealing with this case. Motivated by Theorem 6.2 one could attempt to solve the homotopy problem by minimizing E in the homotopy class of u0 . However, while H 1,2 (Σ; N ) is

6. Harmonic Maps of Riemannian Surfaces

237

weakly closed in the topology of H 1,2 (Σ; IRn ), in general this will not be the case for homotopy classes of non-constant maps. Consider for example the family (ϕε )ε>0 of stereographic projections to the standard sphere, introduced in (5.5); projecting back with ϕ1 , we obtain a family of maps uε = ϕε ◦ (ϕ1 )−1 : S 2 → S 2 of degree 1, converging weakly to a constant map. Therefore, the direct method fails to be applicable for solving the homotopy problem. Also note that the space H 1,2 (Σ; N ) is not a manifold; therefore the standard deformation lemma Theorem II.3.4 or II.3.11 cannot be applied. (An interesting variant was recently proposed by Duc [2].) Existence and Non-existence Results In fact, the inﬁnum of E in a given homotopy class in general need not be attained. (See White [1] for further results in this regard.) Even worse, the homotopy problem may not always have a solution. The following result, analogous to the results of Pohoˇzaev (Theorem 1.3) and Wente (Theorem 5.1) in our previous examples, is due to Eells-Wood [1]. 6.3 Theorem. Any harmonic map u ∈ C 1 (T 2 ; S 2 ) necessarily has topological degree = 1. In particular, there is no harmonic map homotopic to a map u0 : T 2 → S 2 of degree +1. This result shows that we may encounter some lack of compactness in attempting to ﬁnd critical points of E. However, compactness can be restored under certain conditions. In a pioneering paper Eells-Sampson [1] have obtained the following result. 6.4 Theorem. Suppose the sectional curvature κN of N is non-positive. Then for any map u0 : Σ → N there exists a harmonic map homotopic to u0 . Surprisingly, also a topological condition on the target may suﬃce to solve the homotopy problem. The following result was obtained independently by Lemaire [1] and Sacks-Uhlenbeck [1]. 6.5 Theorem. If π2 (N ) = 0, then for any u0 ∈ H 1,2 (Σ; N ) there is a smooth harmonic map homotopic to u0 . For the proof of Theorem 6.4 Eells and Sampson consider an evolution problem similar to the ﬂow approach that we have encountered in Section 4. In fact, also Theorem 6.5 can be obtained from an in-depth analysis of the Eells-Sampson ﬂow. Moreover, as we shall see, this analysis reveals a deep analogy between the harmonic map problem and the problems (1.1), (1.3), the Yamabe problem, and the Dirichlet problem for the equation of constant mean curvature (5.1).

238

Chapter III. Limit Cases of the Palais-Smale Condition

The Heat Flow for Harmonic Maps Given a map u0 ∈ C ∞ (Σ; N ), Eells and Sampson propose to consider a solution u: M × [0, ∞[→ N → IRn of the initial value problem (6.1) (6.2)

ut − ΔΣ u = A(u)(∇u, ∇u)Σ ⊥ Tu N in Σ × IR+ , u|t=0 = u0 .

We write ut = ∂t u for brevity. Equation (6.1) may be interpreted as the “L2 gradient ﬂow” for E; in particular, we have the energy inequality T |ut |2 dμ dt + E u(T ) ≤ E(u0 ) , 0

Σ

for all T > 0; see Lemma 6.8 below. Moreover, like the standard linear heat equation, the ﬂow deﬁned by (6.1) has certain smoothing properties. For example, analogous to the stationary (time-independent) case, for (6.1) there holds the Bochner-type inequality (∂t − ΔΣ )e(u) + |∇du|2Σ ≤ C(1 + e(u))e(u) . Working intrinsically, as in Jost [3], formula (3.2.11), the bound for the leading term on the right can be improved and we obtain the diﬀerential inequality (6.3)

(∂t − ΔΣ )e(u) ≤ κN e(u)2 + Ce(u) ,

where κN ≥ 0 again denotes an upper bound for the sectional curvature of N . Proof of Theorem 6.4. If κN ≤ 0, estimate (6.3) implies a linear diﬀerential inequality for the energy density, and we obtain the existence of a global solution u ∈ C 2 (Σ × IR+ , N ) to the evolution problem (6.1), (6.2). Moreover, by the weak Harnack inequality for sub-solutions of parabolic equations (see Moser [2; Theorem 3]) and the energy inequality, the maximum of |∇u| may be a-priori bounded in terms of the initial data. Again by the energy inequality, we can ﬁnd a sequence of times tm → ∞ such that ut (tm ) → 0 in L2 as m → ∞, and a subsequence (u(tm )) converges to a harmonic map homotopic to u0 . By Theorem 6.3, in general we cannot expect that (6.1), (6.2) admits a global smooth solution, converging asymptotically to a harmonic map. However, we propose to establish that (6.1), (6.2) always admits a global weak solution which is unique in its class, for arbitrary initial data u ∈ H 1,2 (Σ; N ) and without imposing any topological or geometric restrictions on the target manifold. Moreover, space-time singularities of the ﬂow or failure of u(t) to converge as t → ∞ may be uniquely associated with the “bubbling oﬀ” of harmonic spheres, as we will now make precise.

6. Harmonic Maps of Riemannian Surfaces

239

The Global Existence Result Let expx : Tx Σ → Σ denote the exponential map at a point x ∈ Σ. (If Σ = T 2 , then expx (y) = x + y.) The following result is due to Struwe [10]. 6.6 Theorem. For any u0 ∈ H 1,2 (Σ; N ) there exists a distribution solution u: Σ × IR+ → N of (6.1) which is smooth on Σ × IR+ away from at most ¯ which satisﬁes the ﬁnitely many points (¯ xk , t¯k ), 0 < t¯k < ∞, 1 ≤ k ≤ K, energy inequality E(u(t)) ≤ E(u(s)) for all 0 ≤ s ≤ t, and which assumes its initial data continuously in H 1,2 (Σ; N ). The solution u is unique in this class. At a singularity (¯ x, t¯) a smooth harmonic map u ¯: S 2 → N separates in the ¯, tm & t¯, Rm 0 as m → ∞ the family sense that for sequences xm → x 2,2 ˜ in Hloc (IR2 ; N ) , um (x) ≡ u expxm (Rm x), tm → u where u ˜: IR2 → N is smooth and harmonic with ﬁnite energy. By composition with stereographic projection, u ˜ induces a non-constant, smooth harmonic map u ¯: S 2 ∼ = IR2 → N . See Fig. 6.1 for illustration. As tm → ∞ suitably, the sequence of maps u(tm ) converges weakly in H 1,2 (Σ; N ) to a smooth harmonic map u∞ : Σ → N , and smoothly away from ˜ where again harmonic spheres separate in ﬁnitely many points x ˜k , 1 ≤ k ≤ K, the sense described above. Moreover, we have E(u∞ ) ≤ E(u0 ) − K ε0 , ¯ +K ˜ is the total number of singularities and where where K = K ε0 = inf E(u) ; u ∈ C 1 (S 2 ; N ) is non-constant and harmonic > 0 is a constant depending only on the geometry of N . In particular, the number of singularities of u is a-priori bounded, K ≤ ε−1 0 E(u0 ). Theorem 6.6 implies Theorem 6.5: Proof of Theorem 6.5. We follow Struwe [15; p. 299 f.]. Let [u0 ] be a homotopy class of maps from Σ into N . We may suppose that [u0 ] is represented by a smooth map u0 : Σ → N such that E(u0 ) ≤ inf E(u) + u∈[u0 ]

ε0 . 2

Let u: Σ × IR+ → N be the solution to the evolution problem (6.1), (6.2) constructed in Theorem 6.6. Suppose u ﬁrst becomes singular or concentrates ¯, rm 0 we have at a point (¯ x, t¯), t¯ ≤ ∞. Then for sequences tm & t¯, xm → x 2,2 (IR2 ; N ) ˜ in Hloc um (x) := u expxm (rm x), tm → u where u ˜ may be extended to a smooth harmonic map u ¯: S 2 → N .

240

Chapter III. Limit Cases of the Palais-Smale Condition

Fig. 6.1. “Separation of spheres”

In the image covered by um we now replace a large part of the “harmonic sphere” u ¯ by its “small” complement, “saving” at least ε0 /2 in energy. If π2 (N ) = 0, this change of um will not aﬀect the homotopy class [um ] = [u0 ], and a contradiction will result. 2,2 ˜ in Hloc and since E(˜ u) < ∞, we can ﬁnd a More precisely, since um → u sequence of radii Rm → ∞ such that as m → ∞ we have rm Rm → 0 while ˜ | + Rm |∇(um − u ˜)|2 do → 0 , sup |um − u ∂BRm

∂BRm

IR2 \BRm

|∇˜ u|2 dx → 0 ,

|∇(um − u ˜)|2 dx → 0 , BRm

where BRm = BRm (0). Now let ψ ∈ H 1,2 (BRm ; IRn ) solve the Dirichlet problem −Δψm = 0 in BRm with boundary data ˜ ψ m = um − u

on ∂BRm .

By the maximum principle ˜| → 0 . sup |ψm | = sup |um − u BRm

∂BRm

6. Harmonic Maps of Riemannian Surfaces

241

Moreover, by classical potential estimates and interpolation (see Lions-Magenes [1; Theorem 8.2]), the Dirichlet integral of ψm may be estimated in terms of the semi-norm of the trace ψm |∂BRm ∈ H 1/2,2 (∂BRm ) and as m → ∞ we obtain |∇ψm |2 dx ≤ c|ψm |2H 1/2,2 (∂BR ) BRm

m

≤c ∂BRm

≤ c Rm

|ψm (x) − ψm (y)|2 dx dy |x − y|2 ∂BRm |∇ ψm |2 do ≤ c Rm |∇(um − u ˜)|2 do → 0 ,

∂BRm

∂BRm

where ∇ denotes the tangential gradient. (See for instance Adams [1; Theorem 7.48] for the equivalent integral representation of the H 1/2,2 -semi-norm.) Hence if we replace u(tm ) by the map vm , where 5 |x| ≥ Rm um (x), 2 x vm expxm (rm x) = u ˜ Rm (x), |x| < Rm , + ψ m |x|2 is deﬁned via the exponential map in a small coordinate chart U around xm , and where vm ≡ u(tm ) on Σ \ U, we obtain that, as m → ∞, dist vm (x), N ≤ sup |ψm | → 0 , BRm

uniformly in x ∈ Σ. But then we may project vm onto N to obtain a map wm ∈ H 1,2 (Σ; N ) satisfying, with error o(1) → 0 as m → ∞, E(wm ) ≤ E(vm ) + o(1) 1 2 ≤ E(um ) − |∇um | dx + |∇˜ u|2 dx 2 BRm IR2 \BRm + |∇ψm |2 dx + o(1) BRm

1 ≤ E(u0 ) − 2

|∇˜ u|2 dx + o(1) BRm

≤ inf E(u) + ε0 /2 − ε0 + o(1) u∈[u0 ]

and the latter is strictly smaller than inf u∈[u0 ] E(u) for large m. On the other hand, since π2 (N ) = 0, the maps x → u ˜(x) and x → 2 x/|x|2 ) for any m are homotopic as maps from BRm into N with ﬁxed u ˜(Rm boundary; hence wm is homotopic to u(tm ), and the latter is homotopic to u0 via the ﬂow (u(t))0≤t≤tm , which is regular for all m by assumption. Thus, if the ﬂow (u(t))t>0 were to develop a singularity at some ﬁnite time t¯ or a concentration as t → ∞, there results a contradiction to our choice of u0 . Hence, the ﬂow (u(t))t>0 exists and is regular for all time and, as tm → ∞ suitably,

242

Chapter III. Limit Cases of the Palais-Smale Condition

u(tm ) → u∞ strongly in H 1,2 (Σ; N ) where u∞ is harmonic. But by Theorem 6.2, for suﬃciently large m the maps u∞ and u(tm ) are homotopic, and the latter is homotopic to u0 via the ﬂow (u(t))0≤t≤tm . The Proof of Theorem 6.6 The following Sobolev interpolation estimate independently due to GagliardiNirenberg and Ladyzhenskaya [1] plays a key role in the proof of Theorem 6.6. 1,2 (IR2 ), any R > 0, and any function 6.7 Lemma. For any function u ∈ Hloc ∞ ϕ ∈ C0 (BR ) with 0 ≤ ϕ ≤ 1 and such that |∇ϕ| ≤ 4/R, there holds |u|4 ϕ2 dx ≤ c0 |u|2 dx · |∇u|2 ϕ2 dx + R−2 |u|2 dx , IR2

BR

BR

BR

with c0 independent of u and R. Proof. The function |u|2 ϕ has compact support. Thus, for any x = (ξ, η) ∈ IR2 we have ξ ∂ 2 2 |u| ϕ (ξ, η) = |u| ϕ (ξ , η) dξ −∞ ∂ξ ∞ ≤2 |∇u| |u|ϕ + |u|2 |∇ϕ| (ξ , η) dξ , −∞

and an analogous estimate with integration in η-direction. Hence by Fubini’s theorem and H¨ older’s inequality we obtain ∞ ∞ |u|4 ϕ2 (x) dx = |u|4 ϕ2 (ξ, η) dξ dη −∞ −∞ IR2 ∞ ∞ ∞ 2 ≤4· ∇u| |u|ϕ + |u| |∇ϕ| (ξ , η) dξ · −∞ −∞ −∞ ∞ dξ dη |∇u| |u|ϕ + |u|2 |∇ϕ| (ξ, η ) dη · −∞ ∞ ∞ =4 |∇u| |u|ϕ + |u|2 |∇ϕ| (ξ , η) dξ dη· −∞ −∞ ∞ ∞ |∇u| |u|ϕ + |u|2 |∇ϕ| (ξ, η )dξ dη · −∞

=4 IR ≤8

2

−∞

2

|∇u| |u|ϕ + |u| |∇ϕ| dx |∇u|2 ϕ2 + |u|2 |∇ϕ|2 dx . |u|2 dx ·

supp ϕ

2

IR2

Since supp ϕ ⊂ BR and |∇ϕ| ≤ 4/R, the claim follows.

6. Harmonic Maps of Riemannian Surfaces

243

To see how we may apply Lemma 6.7, observe that by compactness of N with a uniform constant C = C(N ) for any solution u of (6.1) there holds |ut − ΔΣ u| ≤ C|∇u|2 .

(6.4)

Hence, as shown in Lemmas 6.10 and 6.11 below, if the energy of u(t) is uniformly small on small balls, from Lemma 6.7 applied to the function |∇u| we obtain an L2 -estimate for |∇2 u|, and then also higher regularity. It is then of crucial importance to obtain energy bounds for u(t), both globally and locally. First we note the following energy inequality. 6.8 Lemma. Let u ∈ C 2 Σ × [0, T [; N solve (6.1), (6.2). Then there holds t |ut |2 dμ dt + E u(t) ≤ E(u0 ) for all t < T . 0

Σ

Proof. Multiply (6.1) by ut ∈ Tu N and use orthogonality A(u)(·, ·) ⊥ Tu N to obtain the identity (6.5)

ut − ΔΣ u, ut = |ut |2 +

d e(u) − div ∇u, ut = 0 . dt

The claim follows upon integrating over Σ × [0, t]. Lemma 6.7 also calls for control of the energy density, locally. For this let ıΣ be the injectivity radius of the exponential map on Σ. Then on any ball BR (x0 ) ⊂ Σ of radius R < ıΣ we may introduce Euclidean coordinates by a conformal change of variables. Recall that E is conformally invariant. With reference to such a conformal chart then introduce the local energy 1 |∇u|2 dx , R < ıΣ . E u; BR (x0 ) = 2 BR (x0 ) In the following we shift x0 = 0 and again let BR = BR (0) for brevity. 1 6.9 Lemma. There is an absolute constant c1 such that for any R < 2 ıΣ and 2 any solution u ∈ C (B2R × [0, T ]; N ) to (6.1) with sup0≤t≤T E u(t); B2R ≤ E0 there holds T E u(T ); BR ≤ E(u0 ; B2R ) + c1 2 E0 . R

Proof. Consider Σ = T 2 , for simplicity. Choose ϕ ∈ C0∞ (B2R ) such that 0 ≤ ϕ ≤ 1, ϕ ≡ 1 in BR and |∇ϕ| ≤ 2/R. Multiplying (6.5) by ϕ2 and integrating by parts, we obtain T T 1 d 2 2 2 2 |∇u| ϕ dx dt = −2 ∇uut ∇ϕϕ dx dt |ut | ϕ + 2 dt 0 0 B2R B2R T T ≤ |ut |2 ϕ2 dx dt + |∇u|2 |∇ϕ|2 dx dt . 0

B2R

0

B2R

244

Chapter III. Limit Cases of the Palais-Smale Condition

Therefore we can estimate T 1 |∇u|2 ϕ2 dxt=0 E u(T ); BR − E(u0 , B2R ) ≤ 2 B2R T T ≤ 4R−2 |∇u|2 dx dt ≤ 8 2 E0 , R 0 B2R as claimed. If the energy of a solution u to (6.1) is uniformly small on small balls, we can bound its second derivatives in terms of the energy. 6.10 Lemma. There exists ε1 = ε1 (Σ, N ) > 0 with the following property: If u ∈ C 2 B2R × [0, T [; N solves (6.1) on B2R × [0, T [ for some R ∈]0, 12 ıΣ [, and if sup E u(t), B2R < ε1 ≤ E0 , 0≤t≤T

then we have

T

T |∇ u| dx dt ≤ c E0 1 + 2 , R BR 2

0

2

where c depends only on Σ and N . Proof. Let Q = B2R ×[0, T [. Choose ϕ ∈ C0∞ (B2R ) such that 0 ≤ ϕ ≤ 1, ϕ ≡ 1 on BR (x), |∇ϕ| ≤ 2/R and multiply (6.1) by Δuϕ2 . Then from (6.4) we obtain that |∇u|2 ϕ2 2 2 + |Δu| ϕ ∂t dx dt 2 Q 2 2 |∇u| |Δu|ϕ dx dt − 2 ut ∇u∇ϕϕ dx dt ≤C Q Q 1 |Δu|2 ϕ2 dx dt |∇u|4 ϕ2 + |∇u|2 |∇ϕ|2 dx dt + ≤C 2 Q Q 1 2 2 2 2 2 |∇ u| ϕ dx dt + |Δu| ϕ dx dt ≤ C1 ε1 2 Q Q |∇u|2 dx dt . + CR−2 Q

Here we have used the fact that ut ∇u = Δu∇u by (6.1). Young’s inequality 2|ab| ≤ δa2 + δ −1 b2 for any a, b ∈ IR, δ > 0 was used to pass from the second to the third line, and we invoked Lemma 6.7 in the ﬁnal estimate. Integrating by parts twice and again using Young’s inequality, we ﬁnd the estimate 1 2 2 2 2 2 |Δu| ϕ dx ≥ |∇ u| ϕ dx − C |∇u|2 |∇ϕ|2 dx . 2 Q Q Q

6. Harmonic Maps of Riemannian Surfaces

245

Hence, if we choose ε1 > 0 such that C1 ε1 ≤ 18 , we obtain

|∇2 u|2 ϕ2 dx dt ≤ 4 Q

|∇u0 |2 ϕ2 dx + CR−2

Σ

T ≤ C E0 1 + 2 . R

T

|∇u|2 dx dt

0

B2R

By choice of ϕ this implies the assertion of the lemma. As ﬁnal preparation we show that if u ∈ H 1,2 weakly solves (6.1) on some interval ]0, T [ with |∇2 u| ∈ L2 , then u is smooth and smoothly extends up to time T . 6.11 Lemma. Let Q = BR ×]0, T [. If u ∈ H 1,2 (Q; N ) solves (6.1) on Q with |∇2 u| ∈ L2 (Q) and with sup0 0. We proceed by Moser For any L ≥ 1 let |∇u|L = iteration. min{|∇u|, L}. Multiply (6.1) by − div ∇u|∇u|sL and integrate by parts at any (ﬁxed) time t < T . Dropping the term s s ∇(|∇u|2 ) · ∇(|∇u|2L )|∇u|s−2 dx = |∇(|∇u|2L )|2 |∇u|s−2 dx ≥ 0 , L L 4 Σ 4 Σ on account of (6.4) for any s ≥ 0 we obtain the estimate |∇u|2 s 2 2 s |∇u|L + |∇ u| |∇u|L dx ≤ C |∇2 u||∇u|sL |∇u|2 dx ∂t 2 Σ Σ 1 ≤ |∇2 u|2 |∇u|sL dx + C |∇u|4 |∇u|sL dx . 2 Σ Σ The ﬁrst term on the right is easily absorbed in the left hand side. Denoting s/2 wL,s = |∇u|L |∇u|, upon integrating in t, for any 0 < t0 < t1 < T we have

t1

t0

∂t |∇u|2 |∇u|sL dx dt =

Σ

|wL,s (t1 )|2 dx −

Σ

t1 s dx dt ∂t |∇u|s+2 L s + 2 t0 Σ 2 |wL,s (t1 )|2 dx − |wL,s (t0 )|2 dx . ≥ s+2 Σ Σ

|wL,s (t0 )|2 dx Σ

−

Thus, with constants depending only on Σ, N , and s we ﬁnd

246

Chapter III. Limit Cases of the Palais-Smale Condition

|wL,s | dx +

sup t0
≤C

Σ t1

≤C

|∇wL,s |2 dx dt

t0

Σ

|∇u| |wL,s | dx dt + C 2

t0

t1

2

Σ t1

t0

|wL,s (t0 )|2 dx

2

|∇u| dx dt · 4

Σ

t1

Σ

t0

|wL,s | dx dt 4

12

|wL,s (t0 )|2 dx .

+C

Σ

Σ

But by Lemma 6.7 for any L ≥ 1, s ≥ 0 on Q0 = Σ × [t0 , t1 ] we can bound wL,s 4L4 (Q0 ) ≤ c0 sup wL,s (t)2L2 (Σ) · ∇wL,s 2L2 (Q0 ) + wL,s 2L2 (Q0 ) . t0 ≤t≤t1

Provided that t1 ≤ t0 + 1, this yields the estimate sup |wL,s |4 dx dt ≤ c0 |wL,s |2 dx + t0
Q0

Σ

|∇wL,s |2 dx dt

Σ

t0

.

Q0

Together with our previous bounds we then obtain t1 t1 |wL,s |4 dx dt ≤ c1 |∇u|4 dx dt · t0

2

Σ

t1

t0

|wL,s |4 dx dt Σ

|wL,s (t0 )| dx 2

+C

2 .

Σ

Since u has ﬁnite energy and |∇2 u| ∈ L2 (Q), Lemma 6.7 in particular yields that |∇u| ∈ L4 (Q). By absolute continuity of the Lebesgue integral we may then choose γ = γ(s, Σ, N ) > 0 such that for 0 ≤ t0 < t1 ≤ T with t1 − t0 < γ there holds t1 |∇u|4 dx dt ≤ 1 . 2c1 t0

Σ

For any L ≥ 1, any s ≥ 0, and any 0 < t0 < t1 < t0 + γ, with a constant C depending only on the geometry and s there then results the estimate 2 t1 |wL,s |4 dx dt ≤ C |wL,s (t0 )|2 dx . t0

Σ

Σ

Let s0 = 0, si = 2(1 + si−1 ), i ∈ IN. Given t∗ > 0, suppose that |∇u(ti−1 )| ∈ L2+si−1 (Σ) for some i ∈ IN, where 0 = t0 ≤ ti−1 < t∗ . We may assume that t∗ < min{ti−1 + γ, T }. Upon letting L → ∞ then we can bound 2 t∗ 2+si 2+si−1 |∇u| dx dt ≤ C |∇u(ti−1 )| dx . ti−1

Σ

Σ

Covering the interval ]ti−1 , T [ with ﬁnitely many intervals of length γ, we ﬁnd that ∇u ∈ L2+si Σ × [t∗ , T ] . Moreover, by Fubini’s theorem, there exists ti ∈ ]ti−1 , t∗ [ such that |∇u(ti )| ∈ L2+si (Σ). Induction then yields that |∇u(ti )| ∈

6. Harmonic Maps of Riemannian Surfaces

247

i L2+s 0 ≤ ti < t∗ for any i ∈ IN, and we conclude that ∇u ∈ (Σ) for suitable Lq Σ × [t∗ , T ] for any q < ∞ and any t∗ > 0, as claimed. The linear theory for (6.4) then yields that |ut |, |∇2 u| ∈ Lq Σ × [t∗ , T ] for any q < ∞, and ∇u is uniformly H¨ older continuous on Σ × [t∗ , T ], for any t∗ > 0; see Ladyzhenskaya-Solonnikov-Ural’ceva [1; Theorem IV.9.1 and Lemma II.3.3]. Higher regularity up to t = T then follows from the Schauder estimates for the heat equation, applied to (6.1), and the usual bootstrap argument.

Proof of Theorem 6.6. For simplicity, we always consider the case Σ = T 2 . (1◦ ) We ﬁrst consider smooth initial data u0 ∈ C ∞ (Σ; N ). By the local solvability of ordinary diﬀerential equations in Banach spaces, problem (6.1), (6.2) has a local solution u ∈ C ∞ (Σ × [0, T [; N ); see Hamilton [1; p. 122 ﬀ.]. By Lemma 6.8 we have ut ∈ L2 (Σ × [0, T ]) and E u(t) ≤ E u(s) ≤ E(u0 ) for all 0 ≤ s ≤ t ≤ T . Moreover, if R > 0 can be chosen such that E u(t); BR(x) < ε1 , sup x∈Σ, 0≤t≤T

it follows from Lemma 6.10 that T T 2 2 |∇ u| dx dt ≤ c E(u0 ) 1 + 2 , R 0 Σ and hence from Lemma 6.11 that u extends to a C ∞ -solution of (6.1) on the closure Σ × [0, T ]. But then u can be extended beyond T as a smooth solution of (6.1). Thus, if T > 0 is maximal, there exist points x1 , x2 , . . . such that lim sup E u(t), BR (xk ) ≥ ε1 tT

for any R > 0 and any index k. Choose any ﬁnite collection xk , 1 ≤ k ≤ K, of such points and for any R > 0, k = 1, . . . , K, let tk < T be chosen such that E u(tk ); BR(xk ) ≥ ε1 /2 . 2

ε1 R =: t0 ≥ 0 and that B2R (xk ) ∩ B2R (xj ) = ∅ We may assume that tk ≥ T − 4c 1 E0 for j = k. With the constant c1 from Lemma 6.9 then we have K E u(t0 ); B2R (xk ) E u(t0 ) ≥ k=1

K tk − t0 E E u(tk ); BR (xk ) − c1 0 R2 k=1 T − t0 ≥ K ε1 /2 − c1 E0 = Kε1 /4 . R2 Since by Lemma 6.8 we have E u(t0 ) ≤ E(u0 ), this gives an upper bound

(6.6)

≥

248

Chapter III. Limit Cases of the Palais-Smale Condition

K ≤ 4E(u0 )/ε1 for the number K of singular points x1 , . . . , xK at time T . Moreover for any Q ⊂⊂ Σ × [0, T ] \ {(x1 , T ), . . . , (xK , T )} there exists R = RQ > 0 such that sup E u(t); BR (x) < ε1 (x,t)∈Q

and by Lemma 6.10, 6.11 our solution u extends to a C ∞ -solution of (6.1) on Σ × [0, T ] \ {(x1 , T ), . . . (xK , T )}. (2◦) For initial data u0 ∈ H 1,2 (Σ; N ) choose a sequence u0m ∈ C ∞ (Σ; N ) approximating u0 in H 1,2 (Σ; N ). This is possible by Theorem 6.2. For each m let um be the associated solution of (6.1) with um (0) = u0m and Tm > 0 its maximal time of existence. Let E0 = supm E(u0m ). Choose R0 > 0 such that sup E u0 ; B2R0 (x) ≤ ε1 /4 . x∈Σ

This inequality will also hold with ε1 /2 instead of ε1 /4 for u0m if m ≥ m0 . By ε R2 Lemma 6.9 then for T = 4c11 E00 we have sup x∈Σ 0≤t≤min{Tm ,T }

T E um (t); BR0 (x) ≤ sup E u0m , B2R0 (x) + c1 2 E0 R x∈Σ 0 ≤ ε1 /2 + ε1 /4 < ε1 .

By Lemma 6.10 it follows that ∇2 um is uniformly bounded in L2 Σ × [0, t] for t ≤ min{T, Tm } in terms of E(u0 ) and R0 only. By Lemma 6.11, therefore, the interval of existence of um is both open and closed in [0, T ]; that is, Tm ≥ T > 0. Moreover T T 2 2 |∇ um | dx dt ≤ c E0 1 + 2 R0 0 Σ weakly uniformly. Also using Lemma 6.8, we may assume that um converges to a solution u of (6.1), (6.2) with |ut |, |∇2 u|2 ∈ L2 Σ × [0, T ] , and such that E u(t) ≤ lim inf E um (t) ≤ lim E(u0m ) = E(u0 ) , m→∞

m→∞

2

uniformly in t ∈ [0, T ]. Since |ut | ∈ L Σ × [0, T ] , the solution u attains 2 its initial data u0 continuously in L (Σ; N ); by the1,2uniform energy bound E u(t) ≤ E(u0 ), moreover, this is also true in the H (Σ; N )-topology. By Lemma 6.11 we have u ∈ C ∞ (Σ×]0, T1 [, N ) for some maximal T1 > T . By Lemma 6.8, we then also have the energy estimate E u(t) ≤ E u(s) ≤ E(u0 ) for 0 ≤ s ≤ t < T1 . By part (1◦ ) of this proof, u extends smoothly to Σ×]0, T1 ] \ {(x1 , T1 ), . . . , (xK1 , T1 )} for some ﬁnite collection of singular points (1) xk , 1 ≤ k ≤ K1 . Moreover, as t & T1 we have u(t) → u0 ∈ H 1,2 (Σ; N ) 1,2 weakly and strongly in Hloc Σ \ {x1 , . . . , xK1 }; N . Thus, by (6.6),

6. Harmonic Maps of Riemannian Surfaces

249

K1 (1) (1) E(u0 ) = lim E u0 ; Σ \ B2R (xk ) R→0

= lim lim

R→0 tT1

k=1 K1 E u(t); B2R (xk ) E u(t); Σ −

k=1

≤ E(u0 ) − K1 ε1 /4 . (0)

Letting T0 = 0, u(0) = u, u0 = u0 , by iteration we now obtain a sequence (m) u(m) of smooth solutions to (6.1) on Σ×]Tm , Tm+1 [ with initial data u0 at (m+1) weakly in H 1,2 (Σ; N ) as t & Tm+1 . t = Tm and such that u(m) (t) → u0 (m) (m) Moreover, u(m) has ﬁnitely many singularities x1 , . . . , xKm+1 at t = Tm+1 , where (6.7)

m+1 (m) (m+1) ≤ E u0 − Km+1 ε1 /4 ≤ . . . ≤ E(u0 ) − Kl ε1 /4 , E u0 l=1 (m+1) u0

(m) (m) x1 , . . . , xKm+1

smoothly away from as t & Tm+1 , for and u (t) → any m ∈ IN. In particular, the total number of singularities of the ﬂows u(m) is ﬁnite, and Tm = ∞ for some m ∈ IN. Piecing the u(m) together, for any initial u u0 ∈ H 1,2 (Σ; N ) we thus obtain a weak solution to (6.1), (6.2) on Σ×]0, ∞[ satisfying the energy inequality E u(t) ≤ E u(s) ≤ E(u0 ) for 0 ≤ s ≤ t < ∞, 1,2 attaining the initial data continuously in H Σ; N , and smooth on Σ×]0, ∞[ up to ﬁnitely many points. (m)

(3◦) Asymptotics: If for some T > 0, R > 0 we have sup E u(t); BR(x) < ε1 , x∈Σ, t>T

then, by Lemma 6.10, for any t > T there holds t+1 |∇2 u|2 dx dt ≤ c E(u0 ) 1 + R−2 t

Σ

with a uniform constant c = c(Σ, N ), while by Lemma 6.8 we have t+1 |ut |2 dx dt → 0 . t

Σ

Hence we may choose a sequence tm → ∞ such that u(tm ) → u∞ weakly in H 2,2 (Σ; N ), while ut (tm ) → 0 in L2 . Moreover, by the Rellich-Kondrakov theorem u(tm ) → u∞ also strongly in W 1,p (Σ; N ) for any p < ∞. Testing (6.1) with Δ(u − u∞ ) and integrating by parts on Σ × {tm }, we obtain that |∇2 (u − u∞ )|2 dx Σ×{tm } ut Δ(u − u∞ )dx + A(u)(∇u, ∇u)Δ(u − u∞ )dx → 0 ; ≤ Σ×{tm }

Σ×{tm }

250

Chapter III. Limit Cases of the Palais-Smale Condition

that is, u(tm ) → u∞ also strongly in H 2,2 (Σ; N ). Passing to the limit m → ∞ in (6.1), evaluated at tm , we then ﬁnd that u∞ is harmonic. It remains to study the case that for a sequence sm → ∞ there exist points x0m ∈ Σ such that for any R > 0 we have lim inf E u(sm ); BR (x0m ) > ε1 /2 . m→∞

Passing to a subsequence, we may assume that x0m → x0 , and there holds (6.8) lim inf E u(sm ); B2R(x0 ) ≥ ε1 /2 m→∞

for any R > 0. Suppose that (6.8) holds for points x1 , . . . , xK . Choose R > 0 such that B2R (xj ) ∩B2R (xk ) = ∅ (j = k). Then for suﬃciently large m we have (6.9)

K E u(sm ) ≥ E u(sm ); B2R (xk ) ≥ Kε1 /4 . k=1

It follows that K ≤ 4E(u0 )/ε1 . We thus may assume that the collection x1 , . . . , xK is maximal with property (6.8). By repeated selection of subsequences of (sm ) and in view of the uniform boundedness of the number K of concentration points (independent of the sequence (sm )), we can even achieve that x1 , . . . , xK is maximal with the property that lim sup E u(sm ); B2R (xk ) ≥ ε1 /2 m→∞

for any R > 0, 1 ≤ k ≤ K. By Lemma 6.9 then for any Ω ⊂⊂ Σ \ {x1 , . . . , xK } R2 there holds there exists R > 0 such that for any x ∈ Ω with τ = 4cε1 1E(u 0) E u(t); BR(x) ≤ lim sup E u(tm ); B2R (x) + ε1 /2 ≤ ε1 sup sm ≤t≤sm +τ

m→∞

for large m. By compactness, Ω is covered by ﬁnitely many such balls BR (x). Hence by Lemma 6.10 there holds sm +τ τ |∇2 u|2 dx dt ≤ c E(u0 ) 1 + 2 = c E(u0 ) + c (6.10) R sm Ω for large m, while Lemma 6.8 implies that sm +τ |ut |2 dx dt → 0 . (6.11) sm

Ω

Exhausting Σ \ {x1 , . . . , xK } by such domains Ω, we may choose a sequence 2,2 (Σ \ {x1 , . . . , xK }; N ), tm ∈ [sm , sm + τ ] such that u(tm ) → u∞ weakly in Hloc where u∞ : Σ \ {x1 , . . . , xK } → N is harmonic. Moreover, by Lemma 6.8 we also may assume that u(tm ) → u∞ weakly in H 1,2 (Σ; N ), and u∞ has ﬁnite energy. By the regularity result of Sacks-Uhlenbeck [1; Theorem 3.6], then u∞ extends to a smooth harmonic map u∞ ∈ C ∞ (Σ; N ).

6. Harmonic Maps of Riemannian Surfaces

251

(4◦) Singularities: Suppose (¯ x, t¯) is singular in the sense that for any R > 0 we have x) ≥ ε1 , lim sup E u(t); BR (¯ tt¯

or that x ¯ is a concentration point at t¯ = ∞ in the sense that for a sequence sm → ∞ as in (3◦ ) for any R > 0 we have lim inf E u(sm ); B2R (¯ x) ≥ ε1 /2 . m→∞

By ﬁniteness of the singular set, x ¯ is isolated among concentration points. ε R2m , Hence there is R0 > 0 such that when we let Rm → 0 and set τm = 16c11E(u 0) there are t¯m & t¯ and x ¯m → x ¯ such that E u(t¯m ); BRm (¯ xm ) = sup E u(t); BRm (x) = ε1 /4 . x∈B2R (¯ x) 0 ¯m −τm ≤t≤t ¯m t

xm ) ⊂ BR0 (¯ x); moreover, if t¯ = ∞ we may suppose that We may assume BRm (¯ sm ≤ t¯m − τm ≤ sm + τm . Rescale 2 um (x, t) := u(¯ xm + Rm x, t¯m + Rm t) 1 , solves (6.1) and note that um : B1+R0 /Rm × [t0 , 0] → N , with t0 = − 16c1εE(u 0) classically with sup E um (t); B1 (x) ≤ E um (0); B1 = ε1 /4 Rm |x|≤R0 t0 ≤t≤0

and

0 t0

|um,t |2 dx dt ≤ BR0 /Rm

t¯m

t¯m −τm

|ut |2 dx dt → 0 , Σ

as m → ∞. From Lemma 6.10 it follows that also 0 |∇2 um |2 dx dt ≤ c t0

BR0 /Rm

uniformly. Hence for a suitable sequence tm ∈ [t0 , 0] we have convergence 2,2 ˜ weakly in Hloc (IR2 ; N ) and strongly um,t (tm ) → 0 in L2 , while um (tm ) → u 1,2 2 ˜ strongly in in Hloc (IR ; N ). (In fact, we can even show that um (tm ) → u 2,2 (IR2 ; N ).) Upon passing to the limit m → ∞ in Equation (6.1) for um at Hloc ˜ is harmonic. Moreover, by Lemma 6.9 with error o(1) → 0 t = tm we see that u as m → ∞ we have E(˜ u; B2 ) = E um (tm ); B2 − o(1) ≥ E um (0); B1 + c1 t0 E(u0 ) − o(1) = ε1 /4 − ε1 /16 − o(1) > 0 for suﬃciently large m. Thus u ˜ ≡ const. Finally,

252

Chapter III. Limit Cases of the Palais-Smale Condition

2 E(˜ u) ≤ lim inf E(um (tm ); BR0 /Rm ) ≤ lim inf E(u(t¯m + Rm tm )) ≤ E(u0 ) , m→∞

m→∞

and by the result of Sacks-Uhlenbeck [1; Theorem 3.6] quoted earlier, u ˜ extends to a harmonic map u ¯: S 2 → N . In particular, we conclude E(¯ u) = E(˜ u) ≥ ε0 . Therefore, and using Lemma 6.9, for any r, R > 0 we obtain that 2 xm ) ≥ E u(t¯m + Rm tm ); BR (¯ xm ) − o(1) E u(¯ sm ); B2R (¯ = E um (tm ); BR/Rm − o(1) ≥ E um (tm ); Br − o(1) =E u ˜m ; Br − o(1) ≥ ε0 − o(1) , where o(1) → 0 as we ﬁrst let m → ∞ and then r → ∞. It follows that estimates (6.6), (6.7), and (6.9) may be improved, yielding the upper bound K ≤ E(u0 )/ε0 for the total number of singularities and concentration points. (5◦) Uniqueness: It suﬃces to show that two solutions u, v of (6.1) satisfying ∂t u , ∇2 u , ∂t v , ∇2 v ∈ L2 Σ × [0, T ] and with u |t=0 = u0 = v |t=0 coincide. Let w = u − v. By (6.1), w satisﬁes the diﬀerential inequality |∂t w − ΔΣ w| = |A(u)(∇u, ∇u) − A(v)(∇v, ∇v)| ≤ C|w| |∇u|2 + |∇v|2 + C|∇w| |∇u| + |∇v| . Multiplying by w and integrating by parts over Σ × [0, t0 ], for any t0 > 0 we obtain t0 1 2 |w(t0 )| dx + |∇w|2 dx dt 2 Σ 0 Σ (6.12) t0 1 t0 2 2 2 |w| |∇u| + |∇v| dx dt + |∇w|2 dx dt . ≤C 2 0 0 Σ Σ Here we used Young’s inequality to estimate 1 C|w||∇w| |∇u| + |∇v| ≤ |∇w|2 + C|w|2 |∇u|2 + |∇v|2 , 2 and we also used the fact that w |t=0 = 0. By Lemma 6.7, the functions |∇u|, |∇v|, and w belong to L4 Σ × [0, t0 ] , and for any δ > 0 we can estimate t0 |∇u|4 + |∇v|4 dx dt 0 Σ t0 2 2 |∇ u| + |∇2 v|2 + |∇u|2 + |∇v|2 dx dt ≤ δ 2 , ≤ CE(u0 ) 0

Σ

if 0 < t0 ≤ T0 = T0 (δ) is suﬃciently small. We may assume that T0 ≤ 1. Then similarly for such t0 we can bound

6. Harmonic Maps of Riemannian Surfaces

t0

253

0

Σ

|w|4 dx dt ≤C sup |w(t)|2 dx · 0≤t≤t0

≤C

Σ

t0

|w(t)|2 dx +

sup 0≤t≤t0

Σ

0

t0

|∇w|2 + |w|2 dx dt

Σ

0

2

|∇w|2 dx dt

.

Σ

Given δ > 0, choose 0 < t0 ≤ T0 = T0 (δ) such that |w(t0 )|2 dx = sup |w(t)|2 dx = sup |w(t)|2 dx. 0≤t≤T0

Σ

0≤t≤t0

Σ

Σ

Then, from (6.12) we obtain the estimate t0 |w(t)|2 dx + |∇w|2 dx dt sup 0≤t≤t0

Σ

0

Σ

1/2

|∇u|4 + |∇v|4 dx dt 0 Σ 0 Σ t0 ≤ C1 δ sup |w(t)|2 dx + |∇w|2 dx dt . ≤C

t0

|w|4 dx dt

0≤t≤t0

Σ

t0

0

1/2

Σ

with a constant C1 = C1 (Σ, N ). Thus, if we choose δ > 0 such that C1 δ < 1, it follows that w ≡ 0 on Σ ×[0, T0 (δ)]. More generally, the above argument shows that the maximal interval I ⊂ [0, T ] containing t = 0 and such that u(t) = v(t) for t ∈ I is relatively open. Since I is trivially closed, uniqueness follows. 6.12 Remark. Chang [6] has obtained the analogue of Theorem 6.6 for the evolution problem (6.1), (6.2) on manifolds Σ with boundary ∂Σ = ∅ and with Dirichlet boundary data. His result – in the same way as we used Theorem 6.6 to prove Theorem 6.5 – can be used to prove, for instance, the existence and multiplicity results of Brezis-Coron [1] and Jost [1] for harmonic maps with boundary. Finite-time Blow-up. The question whether singularities actually may occur in ﬁnite time remained open for quite a while. It was ﬁnally settled by Chang-Ding-Ye [1] who gave the following example. 6.13 The co-rotational setting. Let x = reiϕ be the representation of points x ∈ IR2 in polar coordinates (r, ϕ). Similarly, we denote as θ the longitudinal angle on S 2 and let ρ denote the distance traveled along any ﬁxed geodesic from the north pole N . A map u: B = B1 (0) → S 2 is co-rotational (of degree 1), if u is represented by maps

254

Chapter III. Limit Cases of the Palais-Smale Condition

θ(reiϕ ) = ϕ , ρ(reiϕ ) = h(r) so that u(reiϕ ) = expN (h(r)eiϕ ) . The energy of u then may be expressed as 1 1 sin2 h (6.13) E(u) = |∇u|2 dx = π |hr |2 + dr , 2 B r2 0 and u is harmonic if and only if h satisﬁes the equation sin 2h 1 = 0 on [0, 1[ . − (rhr )r + r 2r 2 Likewise, a smooth family u(t), 0 ≤ t < T , of such maps, represented by a map h: [0, 1] × [0, T [→ IR, evolves by the heat ﬂow for harmonic maps from B to S 2 , that is, solves the equation ut − Δu = u|∇u|2 on B × [0, T [ ,

(6.14)

if and only if h solves the scalar evolution equation (6.15)

1 sin 2h = 0 on [0, 1[×[0, T [ . ht − (rhr )r + r 2r 2

The representation

¯ h(r) = 2 arctan(r)

of the harmonic map u ¯: IR → S 2 obtained as the inverse of stereographic projection then is a stationary (time-independent) solution of (6.15), as are all functions ¯ ¯ λ (r) = h(r/λ) , λ > 0, h 2

¯(x/λ), λ > 0. corresponding to the scaled maps u ¯λ (x) = u Chang-Ding-Ye [1] show the following result. 6.14 Theorem. For any smooth co-rotational data u0 : B → S 2 , represented by a map h0 : [0, 1] → IR with h0 (0) = 0 and h0 (1) = b > π, the solution u to the Equation (6.14) with inital and boundary data u0 , or, equivalently, the solution h to Equation (6.15) with data h(0, r) = h0 (r) , h(t, 0) = h0 (0) = 0 , h(t, 1) = h0 (1) = b , will develop a singularity at a point (0, T ) for some T < ∞ in the sense that |∇u| and |hr | will be unbounded in any neighborhood of the point (0, T ) ∈ B × [0, T ]. Observe that all maps h0 appearing in the Chang-Ding-Ye result cover the sphere slightly more than once. It is an interesting open question what the precise rate of blow-up is.

6. Harmonic Maps of Riemannian Surfaces

255

Proof. Let b > π. We show that (6.15) admits a smooth sub-solution f with f (0, t) = 0 ≤ f (r, t) ≤ f (1, t) ≤ b on [0, 1] × [0, Tf [ for some Tf < ∞ such that fr (0, t) → ∞ as t → Tf . For any data h0 as in the theorem such that h0 ≥ f (·, 0) we then consider the corresponding solution h of (6.15) with initial data h(0) = h0 . Provided that a smooth solution h exists for 0 ≤ t < Tf , by the maximum principle h will satisfy h ≥ f on [0, 1] × [0, Tf [. Since h(0, t) = 0 = f (0, t), it then also follows that hr (0, t) → ∞ as t → Tf . Hence h must blow up at some time T ≤ Tf . For constants ε > 0, μ > 0 consider the function l = l(r), given by ¯ μ (r 1+ε ) . l(r) = h Note that l satisﬁes the equation 1 (1 + ε)2 sin 2l − (rlr )r + = 0. r 2r 2

(6.16)

Fix some number 0 < ε < 1 once and for all. Given b > π as above, we can ﬁx μ > 0 suﬃciently large so that we have l(1) + π ≤ b and so that, in addition, there holds cos l(r) ≥

(6.17)

1 for 0 ≤ r ≤ 1 . 1+ε

With a function λ = λ(t) > 0 to be speciﬁed later we then make the ansatz ¯ λ + l. We have 0 ≤ f (r, t) ≤ b for 0 ≤ r ≤ 1 and all t ≥ 0 where λ(t) is f =h deﬁned. Moreover, in view of (6.15), (6.16) we compute (6.18)

τ (f ) :=

¯ λ(t) − sin 2f + (1 + ε)2 sin 2l sin 2h sin 2f 1 (rfr )r − = . r 2r 2 2r 2

Now observe that for any α, β ∈ IR there holds sin 2(α + β) − sin 2α = sin((2α + β) + β) − sin((2α + β) − β) = 2 cos(2α + β) sin β . Thus we may simplify (6.18) and use (6.17) to obtain ¯ λ(t) + l) sin l τ (f ) = r −2 (1 + ε)2 cos l − cos(2h ¯ λ(t) + l) sin l ≥ r −2 ε sin l . ≥ r −2 1 + ε − cos(2h But for |α| < π/2 we have the identity

256

Chapter III. Limit Cases of the Palais-Smale Condition

sin 2α =

2 tan α . 1 + tan2 α

Since we may assume that μ > 1, it follows that τ (f ) ≥

2μr 1+ε ε ε ε 1+ε sin l = sin(2 arctan(r /μ)) = ≥ ε1 r ε−1 , r2 r2 r 2 μ2 + r 2(1+ε)

where ε1 =

2με μ2 +1

> 0. Given λ0 > 0, δ > 0 deﬁne 1/(1−ε) − (1 − ε)δt λ(t) = λ1−ε 0

so that d λ = −δλε . dt Then we compute 2r d¯ 2δrλε d d . f= h λ= 2 λ(t) = − 2 2 dt dt r + λ dt r + λ2 Hence we obtain 2δλε r d f − τ (f ) ≤ 2 − ε1 r ε−1 = dt r + λ2

2δλε r 2−ε ε−1 − ε . 1 r r 2 + λ2

But by Young’s inequality we have λε r 2−ε ≤ C(ε)(r 2 + λ2 ) . Hence, for any ε > 0 there is δ0 > 0 such that for any 0 < δ < δ0 the function f is a sub-solution to (6.15), as desired. Observe that the constants μ and λ0 in the above construction may be cho0 (0) > 0 sen arbitrarily large. Thus, for any h0 with h0 (1) > π and such that dh dr while h0 (r) > 0 for r > 0 we can arrange that f (·, 0) ≤ h0 and conclude that the solution h to (6.15) must blow up in ﬁnite time. By the maximum principle the above conditions will hold true at any time t > 0 for the solution h(t) issuing from any function h0 ≥ 0 which satisﬁes the hypotheses of the theorem. Finally, observe that for any k ∈ IN the function f − kπ again is a subsolution to (6.15). By iteratively employing the sub-solutions f − kπ for k = k0 , . . . , 1, where k0 is determined so that k0 π +h0 ≥ 0, for a general function h0 as in the theorem we then see that after some waiting-time t0 > 0 the condition h(t) ≥ 0 will hold true for all t ≥ t0 , provided the solution h does not blow up before t0 . The proof is complete.

6. Harmonic Maps of Riemannian Surfaces

257

Reverse Bubbling and Nonuniqueness The fact that the energy E(u(t)) of the solutions u constructed in Theorem 6.6 to the heat ﬂow (6.1), (6.2) is non-increasing in t is essential for the uniqueness of these solutions. In fact, and quite surprisingly, prompted by the work of Chang-Ding-Ye [1] presented above, Topping [1] and Bertsch et al. [1] independently constructed examples of weak solutions to (6.1), (6.2) which violate the energy inequality and therefore diﬀer from the solutions constructed in Theorem 6.6. These solutions, moreover, are smooth with isolated singularities and have bounded energy. The construction intuitively is achieved by attaching more and more concentrated “bubbles” to some smooth given data u0 at time t = 1, and letting the new conﬁguration evolve by the ﬂow, say, for time 1 < t < 2. Under suitable hypotheses on u0 , the resulting sequence of solutions to (6.1) then can be shown to converge to a limit u which blows up as we go backwards to the initial time t = 1 where a “bubble” develops. This process can therefore best be termed “reverse bubbling”. The following result is taken from Topping [1], Theorem 1.1. We again work in the co-rotational setting described in Section 6.12. Recall the representation ¯ h(r) = 2 arctan(r) of the harmonic map obtained as the inverse of stereographic projection and the notation ¯ ¯ λ (r) = h(r/λ), λ>0 h ¯ for the rescaled map h. 6.15 Theorem. Let u0 : B → S 2 be the co-rotational harmonic map represented by the function h0 (r) = π − h(r), thus precisely covering the lower hemisphere. 1 (B ×[0, ∞[; S 2) to (6.1) such There exists a co-rotational weak solution u ∈ Hloc that u(t) = u0 for t ≤ 1 and u(t) = u0 for t > 1. Moreover, u is smooth except at the point (0, 1) ∈ B × [0, ∞[, and there holds the uniform energy bound (6.19)

2π = E(u0 ) ≤ E(u(t)) ≤ E(u0 ) + 4π = lim E(u(t)) for all t . t↓1

Proof. Similar to the proof of Theorem 6.14 we ﬁrst construct a super-solution g to the ﬂow (6.15) which exhibits reverse bubbling at the time t = 0. Many details of the construction may be borrowed from the preceding section. For constants ε > 0, μ > 0 again consider the function l = l(r), given by ¯ μ (r 1+ε ) . l(r) = h Recall that l satisﬁes Equation (6.16), that is, (1 + ε)2 sin 2l 1 = 0. − (rlr )r + r 2r 2

258

Chapter III. Limit Cases of the Palais-Smale Condition

We ﬁx 0 < ε < 1 and then choose μ > 0 suﬃciently large so that the estimates l(1) < π/4 and (6.17) both hold, that is, cos l(r) ≥

1 for 0 ≤ r ≤ 1 . 1+ε

For suitable δ > 0, to be determined in the sequel, we then deﬁne 1/(1−ε) , λ(t) = (1 − ε)δt satisfying d λ = δλε , λ(0) = 0. dt We then make the ansatz ¯λ − l . g=h In the identity sin 2(α + β) − sin 2α = 2 cos(2α + β) sin β , ¯ λ(t) , β = −l. By means of (6.16), (6.17) we then compute we now let α = h ¯ λ(t) − sin 2g − (1 + ε)2 sin 2l sin 2h sin 2g 1 (rgr )r − = 2 r 2r 2 2r −2 2 ¯ =r cos(2hλ(t) − l) − (1 + ε) cos l sin l ¯ λ(t) − l) − (1 + ε) sin l ≤ r −2 cos(2h

τ (g) =

≤− where ε1 =

2με μ2 +1

2εμr ε−1 ε sin l = − ≤ −ε1 r ε−1 , r2 μ2 + r 2(1+ε)

> 0. Moreover, we have d¯ 2δrλε d 2r d g= h λ=− 2 . λ(t) = − 2 2 dt dt r + λ dt r + λ2

Hence we obtain d 2δλε r g − τ (g) ≥ ε1 r ε−1 − 2 = dt r + λ2

ε1 −

2δλε r 2−ε ε−1 ≥ 0, r r 2 + λ2

provided we choose 0 < δ < δ0 suﬃciently small. Next, for n ∈ IN consider the co-rotational solutions un to the ﬂow (6.14) with initial data u0n represented by the functions h(nr)} , 0 ≤ r ≤ 1 . h0n (r) = min{h0 (r), ¯ Note that there is a unique point 0 < rn < 1, tending monotonically to 0 as ¯ for r < rn and h0n (r) = h0 (r) for r > rn ; in n → ∞, such that h0n (r) = h(nr) ¯ = 0 and h0n (1) = h0 (1) = π/2 for all particular, we have that h0n (0) = h(0)

6. Harmonic Maps of Riemannian Surfaces

259

n. Moreover, there holds h0m (r) ≤ h0n (r) for all r ∈ [0, 1], whenever m ≤ n. Finally, from the representation (6.13) we also conclude the bound u, Bnrn ) < E(u0 ) + E(¯ u) = 6π = lim E(u0n ) . E(u0n ) = E(u0 , B \ Brn ) + E(¯ n→∞

Here, in addition, we use the fact that u0 is conformal; its Dirichlet energy therefore equals the area 2π covered by its image. By the energy inequality, Lemma 6.8, we then have u) = 6π E(un (t)) < E(u0 ) + E(¯

(6.20)

for all t > 0. In particular, the ﬂow solutions un will be smooth for t > 0, as any singularity, say at a time tn > 0, by Theorem 6.6 would absorb at least the energy 4π and therefore not leave enough energy for the weak limit un (tn ) to span the boundary data given by u0 , which would require at least energy 2π. Letting hn be the solution to (6.15) representing un , we thus may invoke the maximum principle to conclude that (6.21)

0 ≤ hm (r, t) ≤ hn (r, t) ≤ min{h0 (r), g(r, t)}

for all r ∈ [0, 1], t ≥ 0, whenever m ≤ n. Indeed, since h0 is a time-independent solution of (6.15) and since we have h0m ≤ h0n ≤ h0 for m ≤ n, the ﬁrst inequalities are immediate. Moreover, from our assumption that l(1) < π/4 we deduce that h0 (r) < π − l(r) = limt↓0 g(r, t) for any r > 0. Thus for any n ∈ IN there exists a number τ > 0 such that h0n ≤ g(τ ) and hence hn (t) ≤ g(τ + t) for all t ≥ 0. But by construction the function g(t) is non-increasing in t, and (6.21) follows. In view of the unifom energy bound (6.20) and the pointwise bounds (6.21) a subsequence hn thus converges smoothly away from r = 0 to a solution h of (6.15). We can strengthen this assertion to include the axis r = 0 by means of Lemma 6.10 once we show the following result. 6.16 Lemma. There is a time T > 0 and an index N ∈ IN such that for any t0 ∈]0, T [ there exists a radius r0 > 0 so that sup E(un (t), Br0 (0)) < ε1 for all n ≥ N ,

t0 ≤t≤T

where ε1 > 0 is the constant determined in Lemma 6.10 for N = S 2 . ε1 Proof. We may of course assume that 4π < π2 . Since sup0≤r≤1 h0n (r) → π as n → ∞, we can ﬁnd an index N and a radius r1 := rN > 0 such that ; 1 ε1 h0N (r1 ) = sup h0N (r) ≥ π − . 2 4π 0≤r≤1 We may then ﬁx T > 0 suﬃciently small to insure the inequality

260

Chapter III. Limit Cases of the Palais-Smale Condition

; hN (r1 , t) ≥ π −

ε1 4π

for any t ∈ [0, T ]. By (6.21) we then also have ; ε1 (6.22) hn (r1 , t) ≥ π − 4π for any n ≥ N , uniformly in t ∈ [0, T ]. Reducing T , if necessary, we also may assume that there holds g(1, t) > π/2 for 0 ≤ t ≤ T . For any t0 ∈]0, T [ now we choose 0 < r0 < 1 so that ; ε1 g(r, t0 ) < 4π for r ∈ [0, r0 ]. Recalling that g(r, t) is monotonically decreasing in t for any r, by (6.21) then for any r ∈ [0, r0 ], any t ∈ [t0 , T ], and any n ≥ N we have ; ε1 (6.23) hn (r, t) ≤ g(r, t) ≤ g(r, t0 ) < . 4π In particular, in view of (6.22) and our assumption about ε1 then it follows each t ∈ [t0 , T ] the map hn (t), restricted that 0 < r0 < r1 < 1. Since for ε1 ε1 , π − 4π ], we conclude that the map to [r0 , r1 ] thus covers the interval [ 4π (t), restricted to the annulus B \ B covers the sphere with caps of radius u r1 r0 n ε1 removed from each pole. In particular, since the area of a geodesic disk 4π on S 2 of radius r > 0 is less than πr 2 , the area covered by un (t) over this region ε 2 1 = 4π − ε21 , and hence is larger than 4π − 2π 4π E(un (t), Br1 \ Br0 ) ≥ 4π −

ε1 . 2

By a similar reasoning, and since we have hn (1, t) = π/2 for all t and n, we conclude that the map hn (t), restricted to [r1 , 1] covers the interval [π/2, π − ε1 ], and thus the map un (t), restricted to the annulus B \ Br1 covers the 4π ε1 removed from the south pole. It lower hemisphere with a cap of radius 4π follows that we have ; 2 ε1 ε1 , ≥ 2π − E(un (t), B \ Br1 ) ≥ 2π − π 4π 4 leaving the energy E(un (t), Br0 ) = E(un (t)) − E(un (t), B \ Br1 ) + E(un (t), Br1 \ Br0 ) 3ε1 3ε1 )= < ε1 , ≤ 6π − (6π − 4 4 for the ball Br0 for any t ∈ [t0 , T ], any n ≥ N , as claimed.

6. Harmonic Maps of Riemannian Surfaces

261

Proof of Theorem 6.15, completed. In view of Lemma 6.16, from Lemma 6.10 we obtain uniform bounds for ∇2 un in L2 (B × [t0 , T ]) for any t0 > 0. Together with the smooth estimates on B×]0, T ] away from x = 0 these estimates suﬃce to pass to the limit in Equation (6.14) for un and we obtain a solution u to this equation which splits oﬀ a bubble of area 4π when t decreases to 0. Moreover, u(t) → u0 weakly in H 1 (B) and hence strongly in L2 (B) as t ↓ 0. We may therefore shift time by 1 and join this solution of (6.14) with the constant solution u(t) ≡ u0 for 0 ≤ t ≤ 1 at time t = 1 to obtain a weak solution of (6.14) on the interval [0, 1 + T ] satisfying (6.19). In view of the inequality E(u(t)) ≤ 6π, which by Lemma 6.8 must be strict for t > 1, again the ﬂow u cannot become singular at any time t > 1 and therefore can be smoothly extended for all t > 1. This ﬁnishes the proof. 6.17 Uniqueness issues. Theorem 6.15 also shows that if we drop the requirement that E(u(t)) be non-increasing, the initial value problem for (6.14) with data u0 has more than just the constant solution u(t) ≡ u0 ; in fact, since we may shift time by an arbitrary amount, Theorem 6.15 gives inﬁnitely many distinct solutions that exhibit reverse bubbling at an arbitrary time t0 > 0. On the other hand, results of Rivi`ere [1] and Freire [1], [2] establish uniqueness for the Cauchy problem (6.1), (6.2) also in the “energy class” of weak solutions u to (6.1) such that E(u(t)) is bounded and ∂t u ∈ L2 Σ × [0, T ] , provided that the energy E(u(t)) is a non-increasing function of t. In particular, in view of the smoothing property of the ﬂow (6.1) for short time, their results give a new proof of H´elein’s [1] regularity result for weakly harmonic maps of surfaces. Topping [1] conjectured that uniqueness holds in the energy class if we impose suitable hypotheses to rule out reverse bubbling, for instance, if we suppose that (6.24)

lim sup E(u(t)) < E(u(t0 )) + ε0 , t↓t0

where ε0 is deﬁned in Theorem 6.6. As shown by Rupﬂin [1], this conjecture holds true if, in addition to (6.24), we assume that the function E(u(t)) is of locally bounded variation, or if the constant ε0 in (6.24) is replaced by a possibly smaller constant ε1 > 0, depending only on the target manifold. The preceding remarks refer to arbitrary smooth and closed target manifolds N . In the special case when N = S 2 , from a physical interpretation of the Dirichlet energy E(u, Ω) as the Frank-Oseen energy of a nematic liquid crystal contained in a cylinder of unit length with cross-section Ω, Bertsch et al. [1] conjecture that there is a unique “physical time” t1 ≥ t0 when the energy concentrated in a spherical bubble separating at time t0 from a ﬂow solution u of (6.14) may be “released” as the bubble is re-attached. It may be conjectured that suitable approximations, like the Ginzburg-Landau approximation studied by Harpes [1], can give rise to such uniqueness results for the ﬂow (6.14) in a restricted energy class of solutions that allows both forward and backward bubbling.

262

Chapter III. Limit Cases of the Palais-Smale Condition

6.18 Energy Quantization. At a singular time t¯ < ∞ the ﬂow constructed in Theorem 6.6 splits oﬀ a ﬁnite number of bubbles u ¯j : S 2 → N , j = 1, . . . , k, and we have the estimate (6.25)

limt↑t¯E(u(t)) ≥ E(u(t¯)) +

k

E(¯ uj ) .

j=1

Similar to the quantization result for the Yamabe equation that we obtained in Theorem 3.1, we might expect (6.25) to hold with equality. When N = S n , this was actually shown by Jie Qing [2]. His result was generalized to arbitrary target manifolds by C. Wang [1] and Ding-Tian [1], independently, with later improvements by Lin-Wang [1]. The analogous quantization result (6.26)

lim E(um ) = E(u0 ) +

m→∞

k

E(¯ uj )

j=1

for sequences of harmonic maps um u0 weakly in H 1 (B, N ) with bubbles u ¯j : S 2 → N , j = 1, . . . , k, was shown by Parker [1]. Parker’s result was extended to Palais-Smale sequences (um ) with τ (um ) = Δum − A(um )(∇um , ∇um ) → 0 in L2 and bounded energy by Qing-Tian [1]. By the energy inequality for the heat ﬂow, Lemma 6.8, from (6.26) we then also obtain the corresponding energy quantization in the case when the heat ﬂow concentrates as t → ∞, for suitable sequences tm → ∞. 6.19 Remarks. (1◦ ) Under suitable hypotheses Topping [3] shows exponential L2 -convergence as t → ∞ of the heat ﬂow (6.1) for harmonic maps S 2 → S 2 in the presence of “bubbling at inﬁnity”, and exponential H 1 -convergence locally away from any bubble points of the ﬂow. (2◦) Results analogous to Theorem 6.6 hold for a number of planar geometric ﬂows and may be applied to prove the existence of minimal surfaces or, more generally, of surfaces of constant mean curvature with free boundaries; see Struwe [16]. (3◦) See Schoen-Uhlenbeck [1], [2], Struwe [14], Chen-Struwe [1], Coron [3], Coron-Ghidaglia [1], Chen-Ding [1] for results on harmonic maps and the evolution problem (6.1), (6.2) in the case dim(Σ) > 2. A survey of results related to harmonic maps can be found in Eells-Lemaire [1], [2]. An overview of developments for the evolution problem is given in Struwe [20], [24].

Appendix A

Here, we collect without proof a few basic results about Sobolev spaces. A general reference to this topic is Gilbarg-Trudinger [1], or Adams [1]. Sobolev Spaces 1 Let Ω be a domain in IRn . For nu ∈ Lloc (Ω) and any multi-index α = n (α1 , . . . , αn ) ∈ IN0 , with |α| = j=1 αj , deﬁne the distibutional derivative ∂ α1 ∂ αn α D u = ∂xα1 · · · ∂xαn u by letting n 1 (A.1) < ϕ, Dα u >= (−1)|α| u Dα ϕ dx, Ω

for all ϕ ∈ satisfying

C0∞ (Ω).

We say D u ∈ Lp (Ω), if there is a function gα ∈ Lp (Ω) α ϕgα dx < ϕ, D u >=< ϕ, gα >= α

Ω

for all ϕ ∈ C0∞ (Ω). In this case we identify Dα u with gα ∈ Lp (Ω). For k ∈ IN0 , 1 ≤ p ≤ ∞, deﬁne the space W k,p (Ω) = u ∈ Lp (Ω); Dα u ∈ Lp (Ω) for all α: |α| ≤ k , with norm

upW k,p =

Dα upLp ,

if 1 ≤ p < ∞,

|α|≤k

respectively, with norm uW k,∞ = max Dα uL∞ . |α|≤k

Note that the distributional derivative (A.1) is continuous with respect to weak convergence in L1loc (Ω). Many properties of Lp (Ω) carry over to W k,p (Ω). A.1 Theorem. For any k ∈ IN0 , 1 ≤ p ≤ ∞, W k,p (Ω) is a Banach space. W k,p (Ω) is reﬂexive if and only if 1 < p < ∞. Moreover, W k,2 (Ω) is a Hilbert space with scalar product Dα u Dα v dx , (u, v)W k,2 = |α|≤k

Ω

inducing the norm above. For 1 ≤ p < ∞, W k,p (Ω) also is separable. In fact, we have the following result due to Meyers and Serrin; see Adams [1; Theorem 3.16].

264

Appendix A

A.2 Theorem. For any k ∈ IN0 , 1 ≤ p < ∞, the subspace W k,p ∩ C ∞ (Ω) is dense in W k,p (Ω). The completion of W k,p ∩ C ∞ (Ω) in W k,p (Ω) is denoted by H k,p (Ω). By Theorem A.2, W k,p (Ω) = H k,p (Ω). In particular, if p = 2 it is customary to use the latter notation. Finally, W0k,p (Ω) is the closure of C0∞ (Ω) in W k,p (Ω); in particular, k,2 H0 (Ω) is the closure of C0∞ (Ω) in H k,2 (Ω), with dual H −k (Ω). D k,p (Ω) is the closure of C0∞ (Ω) in the norm Dα upLp . upDk,p = |α|=k

H¨ older Spaces A function u: Ω ⊂ IRn → IR is H¨older continuous with exponent β > 0 if [u](β) = sup x =y∈Ω

|u(x) − u(y)| < ∞. |x − y|β

For m ∈ IN0 , 0 < β ≤ 1, denote C m,β (Ω) = u ∈ C m (Ω); Dα u is H¨older continuous

with exponent β for all α: |α| = m .

¯ becomes a Banach space in the norm If Ω is relatively compact, C m,β (Ω) Dα uL∞ + [Dα u](β) . uC m,β = |α|≤m

|α|=m

The space C m,β (Ω) on an open domain Ω ⊂ IRn carries a Fr´echet space topology, induced by the C m,β -norms on compact sets exhausting Ω. Finally, we may set C m,0 (Ω) := C m (Ω). Observe that for 0 < β ≤ 1 smooth functions are ¯ not dense in C m,β (Ω). Embedding Theorems Let (X, · X ), (Y, · Y ) be Banach spaces. X is (continuously) embedded into Y (denoted X → Y ) if there exists an injective linear map i: X → Y and a constant C such that i(x)Y ≤ CxX , for all x ∈ X. In this case we will often simply identify X with the subspace i(X) ⊂ Y . X is compactly embedded into Y if i maps bounded subsets of X into relatively compact subsets of Y . For the spaces that we are primarily interested in we have the following results. First, from H¨ older’s inequality we obtain:

Appendix A

265

A.3 Theorem. For Ω ⊂ IRn with Lebesgue measure Ln (Ω) < ∞, 1 ≤ p < q ≤ ∞, we have Lq (Ω) → Lp (Ω). This ceases to be true if Ln (Ω) = ∞. For H¨ older spaces, by the theorem of Arz´ela-Ascoli we have the following compactness result; see Adams [1; Theorems 1.30, 1.31]. A.4 Theorem. Suppose Ω is a relatively compact domain in IRn , and let ¯ → C m,α (Ω) ¯ compactly. m ∈ IN0 , 0 ≤ α < β ≤ 1. Then C m,β (Ω) Finally, for Sobolev spaces we have (see Adams [1; Theorem 5.4]): A.5 Theorem (Sobolev embedding theorem). Let Ω ⊂ IRn be a bounded domain with Lipschitz boundary, k ∈ IN, 1 ≤ p ≤ ∞. Then the following holds: np (1◦) If k p < n, we have W k,p (Ω) → Lq (Ω) for 1 ≤ q ≤ n−k p ; the embedding np is compact, if q < n−k p . ¯ for (2◦) If 0 ≤ m < k − np < m + 1, m ∈ IN0 , we have W k,p (Ω) → C m,α (Ω) n n 0 ≤ α ≤ k − m − p ; the embedding is compact, if α < k − m − p . np Compactness of the embedding W k,p (Ω) → Lq (Ω) for q < n−k is a consep quence of the Rellich-Kondrakov theorem; see Adams [1; Theorem 6.2]. Theorem A.5 is valid for W0k,p (Ω)-spaces on arbitrary bounded domains Ω.

Density Theorem By Theorem A.2, Sobolev functions can be approximated by functions enjoying any degree of smoothness in the interior of Ω. Some regularity condition on the boundary ∂Ω is necessary if smoothness up to the boundary is required: A.6 Theorem. Let Ω ⊂ IRn be a bounded domain of class C 1 , and let k ∈ ¯ is dense in W k,p (Ω). IN, 1 ≤ p < ∞. Then C ∞ (Ω) More generally, it suﬃces that Ω has the segment property; see for instance Adams [1; Theorem 3.18]. Trace and Extension Theorems For a domain Ω ⊂⊂ IRn with C k -boundary ∂Ω = Γ , k ∈ IN, 1 < p < ∞, 1 denote as W k− p ,p (Γ ) the space of “traces” u|Γ of functions u ∈ W k,p (Ω). 1 If k = 1 we think of W 1− p ,p (Γ ) as the set of equivalence classes {u} + W01,p (Ω); u ∈ W 1,p (Ω) , endowed with the trace norm = inf vW 1,p (Ω) ; u − v ∈ W01,p (Ω) . u|Γ 1− p1 ,p W

(Γ )

1− p1 ,p

(Γ ) is a Banach space. By this deﬁnition, W In particular, if k = 1 and p = 2, the trace operator u → u|∂Ω is a linear isometry of the (closed) orthogonal complement of H01,2 (Ω) in H 1,2 (Ω) onto 1 H 2 ,2 (Γ ). By the open mapping theorem this provides a bounded “extension 1 operator” H 2 ,2 (Γ ) → H 1,2 (Ω). For general p > 1 (and k = 1) we have:

266

Appendix A

A.7 Theorem. For any Ω ⊂⊂ IRn with C 1 -boundary Γ , 1 < p < ∞ there 1 exists a continuous linear extension operator ext: W 1− p ,p (Γ ) → W 1,p (Ω) such 1 that ext(u) Γ = u, for all u ∈ W 1− p ,p (Γ ). See Adams [1; Theorem 7.53 and 7.55]. Covering ∂Ω = Γ by coordinate patches and deﬁning the Lebesgue space Lp (Γ ) as before via such charts (see Adams [1; 7.51]), an equivalent norm for W s,p (Γ ), where 0 < s < 1, is given by

uW ˜ s,p =

upLp +

Γ

Γ

|u(x) − u(y)|p dx dy |x − y|n−1+sp

1/p ;

see Adams [1; Theorem 7.48]. From this, the following may be deduced: A.8 Theorem. Suppose Ω ⊂⊂ IRn is a domain with C 1 -boundary Γ , 1 < p < 1 ∞. Then W 1,p (Γ ) → W 1− p ,p (Γ ) → Lp (Γ ) and both embeddings are compact. In particular, we have H 1,2 (Ω) → L2 (∂Ω)

(A.2)

compactly, for any bounded domain of class C 1 . Poincar´e Inequality For a bounded domain Ω of diameter d and u ∈ H01,2 (Ω) there holds 2 2 (A.3) |u| dx ≤ d |∇u|2 dx . Ω

Ω

= S, u ∈ C0∞ (Ω) ⊂ C0∞ (S). Then (A.3) It suﬃces to consider Ω ⊂ [0, d]×IR follows immediately from H¨ older’s inequality and the mean value theorem. More generally, we state: n−1

A.9 Theorem. For any bounded domain Ω of class C 1 there exists a constant c = c(Ω) such that for any u ∈ H 1,2 (Ω) we have |u|2 dx ≤ c |∇u|2 dx + c |u|2 do . Ω

Ω

∂Ω

Proof. The argument is modeled on Neˇcas [1; p. 18 f.]. Suppose by contradiction that for a sequence (um ) in H 1,2 (Ω) there holds (A.4) um 2L2 (Ω) ≥ m ∇um 2L2 (Ω) + um 2L2 (∂Ω) .

Appendix A

267

By homogeneity, we may normalize um 2L2 (Ω) = 1. But then (um ) is bounded in H 1,2 (Ω), and we may assume that um → u weakly. Moreover, by Theorem A.5.(1◦ ) , it follows that um → u strongly in L2 (Ω) and by (A.2) also um |∂Ω → u|∂Ω in L2 (∂Ω). But (A.4) also implies that ∇um → 0 in L2 (Ω), and um |∂Ω → 0 in L2 (∂Ω). Hence, u ∈ H01,2 (Ω); moreover, ∇u = 0. By (A.3) therefore, u ≡ 0. But uL2 (Ω) = limm→∞ um L2 (Ω) = 1. Contradiction. In the same spirit the following variant of Poincar´e’s inequality may be derived. A.10 Theorem. Let AR = B2R (0) \ BR (0) ⊂ IRn denote the annulus of size R in IRn . There exists a constant c = c(n, p) such that for any R > 0, any u ∈ H 1,p (AR ) there holds |u − uR |p dx ≤ c Rp |∇u|p dx , AR

AR

where uR denotes the mean of u over the annulus AR . Proof. Scaling with R, we may assume that R = 1, AR = A1 =: A. Moreover, it suﬃces to consider u = u1 = 0. If for a sequence (um ) in H 1,p (A) with um = 0 we have |um |p dx ≥ m

1= A

|∇um |p dx , A

by Theorem A.5 we conclude that um → u ≡ const. = u = 0 in Lp (A). Contradiction.

Appendix B

In this appendix we recall some fundamental estimates for elliptic equations. A basic reference is Gilbarg-Trudinger [1]. On a domain Ω ⊂ IRn we consider second-order elliptic diﬀerential operators of the form ∂2 ∂ u + bi u + cu, (B.1) L u = −aij ∂xj ∂xj ∂xi or in divergence form (B.2)

Lu = −

∂ ∂xi

∂ u +cu , aij ∂xj

with bounded coeﬃcients aij = aji , bi , and c satisfying the ellipticity condition aij ξi ξj ≥ λ|ξ|2 with a uniform constant λ > 0, for all ξ ∈ IRn . By convention, repeated indices are summed from 1 to n. The standard example is the Laplace operator L = −Δ. If aij ∈ C 1 , then any operator of type (B.2) also falls into category ∂ aij . (B.1) with bj = − ∂x i Schauder Estimates Let us ﬁrst consider the (classical) C α -setting; see Gilbarg-Trudinger [1; Theorems 6.2, 6.6]. B.1 Theorem. Let L be an elliptic operator of type (B.1), with coeﬃcients of ¯ Then u ∈ C 2,α (Ω), class C α , and let u ∈ C 2 (Ω). Suppose L u = f ∈ C α (Ω). and for any Ω ⊂⊂ Ω we have (B.3) uC 2,α (Ω ) ≤ C uL∞ (Ω) + f C α (Ω) ¯ . ¯ coincides with a function If in addition Ω is of class C 2+α , and if u ∈ C o (Ω) 2+α ¯ 2,α ¯ (Ω) on ∂Ω, then u ∈ C (Ω) and uo ∈ C ¯ (B.4) uC 2,α (Ω) ¯ ≤ C uL∞ (Ω) + f C α (Ω) ¯ + uo C 2,α (Ω) with constants C possibly depending on L, Ω, n, α, and – in case of (B.3) – on Ω. Lp -theory For solutions in Sobolev spaces the Calder´ on-Zygmund inequality is the counterpart of the Schauder estimates for classical solutions; see Gilbarg-Trudinger [1; Theorems 9.11, 9.13].

Appendix B

269

B.2 Theorem. Let L be elliptic of type (B.1) with continuous coeﬃcients aij . 2,p Suppose u ∈ Wloc (Ω) satisﬁes L u = f in Ω with f ∈ Lp (Ω), 1 < p < ∞. Then for any Ω ⊂⊂ Ω we have uW 2,p (Ω ) ≤ C uLp (Ω) + f Lp (Ω) .

(B.5)

If in addition Ω is of Class C 1,1 , and if there exists a function uo ∈ W 2,p (Ω) such that u − uo ∈ Ho1,p (Ω), then (B.6)

uW 2,p (Ω) ≤ C uLp (Ω) + f Lp (Ω) + uo W 2,p (Ω) .

The constants C may depend on L, Ω, n, p, and – in case of (B.5) – on Ω .

Weak Solutions Let L be elliptic of divergence type (B.2), f ∈ H −1 (Ω). A function u ∈ H01,2 (Ω) weakly solves the equation L u = f if ∂ ∂ u ϕ + c uϕ dx − f ϕdx = 0, for all ϕ ∈ Co∞ (Ω). aij ∂xi ∂xj Ω Ω The integral L(u, ϕ) =

∂ ∂ u ϕ + c uϕ dx aij ∂xi ∂xj Ω

continuously extends to a symmetric bilinear form L on H01,2 (Ω), the Dirichlet form associated with the operator L.

A Regularity Result As an application we consider the equation (B.7)

−Δu = g(·, u) in Ω,

on a domain Ω ⊂ IRn , with a Caratheodory function g: Ω × IR → IR; that is, assuming g(x, u) is measurable in x ∈ Ω and continuous in u ∈ IR. Moreover, we will assume that g satisﬁes the growth condition (B.8)

|g(x, u)| ≤ C 1 + |u|p ,

n+2 where p ≤ n−2 , if n ≥ 3. By (B.8) and Theorem A.5, for any u ∈ H 1,2 (Ω) the composed function g ·, u(·) ∈ H −1 (Ω); see also Theorem C.2. The following estimate is essentially due to Brezis-Kato [1], based on Moser’s [1] iteration technique.

270

Appendix B

B.3 Lemma. Let Ω be a domain in IRn and let g : Ω × IR → IR be a Carath´eodory function such that for almost every x ∈ Ω there holds |g(x, u)| ≤ a(x) 1 + |u| (1◦ ) n/2

1,2 (Ω) be a weak solution with a function 0 ≤ a ∈ Lloc (Ω). Also let u ∈ Hloc q of equation (B.7). Then u ∈ Lloc (Ω) for any q < ∞. If u ∈ H01,2 (Ω), and a ∈ Ln/2 (Ω), then u ∈ Lq (Ω) for any q < ∞.

Proof. Choose η ∈ C0∞ (Ω) and for s ≥ 0, L ≥ 0 let ϕ = ϕs,L = u min |u|2s , L2 η 2 ∈ H01,2 (Ω), with supp ϕ ⊂⊂ Ω. Testing (B.7) with ϕ, we obtain 2 2 2s−2 2 s ∇ |u| |u| |∇u|2 min |u|2s , L2 η 2 dx + η dx 2 {x∈Ω; |u(x)|s ≤L} Ω ∇u u min |u|2s , L2 ∇ηη dx ≤ −2 Ω + a 1 + 2|u|2 min |u|2s , L2 η 2 dx Ω 1 |∇u|2 min |u|2s , L2 η 2 dx + c |u|2 min |u|2s , L2 |∇η|2 dx ≤ 2 Ω Ω 2s 2 2 2 a|u| min |u| , L η dx + |a|η 2 dx . +3 Ω

Ω

Suppose u ∈ L2s+2 (supp(η)). Then for any K ≥ 1 with constants c depending on the L2s+2 -norm of u, restricted to supp(η), there holds ∇ u min |u|s , L η 2 dx ≤ c + c · a|u|2 min |u|2s , L2 η 2 dx Ω Ω 2s 2 2 2 |u| min |u| , L η dx ≤ c + cK Ω a|u|2 min |u|2s , L2 η 2 dx +c {x∈Ω; a(x)≥K}

2/n

≤ c(1 + K) + c ·

a

n/2

dx

{x∈Ω; a(x)≥K}

n−2 n 2n s n−2 × dx u min |u| , L η Ω 2 ≤ c(1 + K) + c1 ε(K) · ∇ u min |u|s , L η dx , Ω

where

2/n

a

ε(K) = {x∈Ω; a(x)≥K}

n/2

dx

→0

(K → ∞) .

Appendix B

271

Fix K such that c1 ε(K) = 12 and observe that for this choice of K (and s as above) we now may conclude that

∇ u min |u|s , L η 2 dx ≤ c(1 + K) Ω

remains uniformly bounded in L. Hence we may let L → ∞ to derive that ∗

|u|s+1 η ∈ H01,2 (Ω) → L2 (Ω) . (2s+2)n

n−2 (Ω). Now That is, whenever u ∈ L2s+2 loc (Ω) we ﬁnd that u ∈ Lloc n iterate, letting s0 = 0, si +1 = (si−1 +1) n−2 , if i ≥ 1, to obtain the conclusion of the lemma. If u ∈ H01,2 (Ω), we may let η = 1 to obtain that u ∈ Lq (Ω) for all q < ∞.

1,2 To apply Lemma B.3, note that, if u ∈ Hloc (Ω) weakly solves (B.7) with a Carath´eodory function g with polynomial growth

|g(x, u)| ≤ C 1 + |u|p , and if p ≤

n+2 n−2

for n > 2, then assumption (1◦ ) of Lemma B.3 is satisﬁed with |g x, u(x) | n/2 a(x) = ∈ Lloc (Ω) . 1 + |u(x)|

By Lemma B.3, therefore, u ∈ Lqloc (Ω), for any q < ∞. In view of our growth condition for g this implies that −Δu = g(u) ∈ Lqloc (Ω) for any q < ∞. Thus, 2,q (Ω), for any by the Cald´eron-Zygmund inequality, Theorem B.2, u ∈ Wloc 1,α q < ∞, whence also u ∈ Cloc (Ω) by the Sobolev embedding theorem, Theorem A.5, for any α < 1. Moreover, if u ∈ H01,2 (Ω), and if ∂Ω ∈ C 2 , by the same token it follows that u ∈ W 2,q ∩ H01,2 (Ω) → C 1,α (Ω). Now we may proceed using Schauder theory. In particular, if g is H¨ older continuous, then u ∈ C 2 (Ω) 2 and is a non-constant, classical C -solution of Equation (B.7). Finally, if g and ∂Ω are smooth, higher regularity (up to the boundary) can be obtained by iterating the Schauder estimates.

Maximum Principle A basic tool for proving existence of solutions to elliptic boundary value problems in H¨older spaces is the maximum principle. We state this in a form due to Walter [1; Theorem 2], allowing for more general coeﬃcients c in the operator L than in classical versions.

272

Appendix B

B.4 Theorem. Suppose L is elliptic of type (B.1) on a domain Ω and suppose ¯ satisﬁes u ∈ C 2 (Ω) ∩ C 1 (Ω) L u ≥ 0 in Ω, and u ≥ 0 on ∂Ω. ¯ such that Moreover, suppose there exists h ∈ C 2 (Ω) ∩ C 0 (Ω) L h ≥ 0 in Ω, and h > 0 on Ω. Then either u > 0 in Ω, or u = βh for some β ≤ 0. ¯ c ∈ In particular, let L be given by (B.2) with coeﬃcients aij ∈ C 1,α (Ω), ¯ Then L is self-adjoint and possesses a complete set of eigenfunctions C α (Ω). ¯ with eigenvalues λ1 < λ2 ≤ λ3 ≤ . . .. Moreover, ϕ1 (ϕj ) in H01,2 (Ω) ∩ C 2,α (Ω) has constant sign, say, ϕ1 > 0 in Ω. Suppose that the ﬁrst Dirichlet eigenvalue λ1 = inf

u =0

(L u, u)L2 ≥ 0. (u, u)L2

Then in Theorem B.4 we may choose h = ϕ1 , and the theorem implies that ¯ of Lu ≥ 0 in Ω, and such that u ≥ 0 on Ω any solution u ∈ C 2 (Ω) ∩ C 1 (Ω) either is positive throughout Ω or vanishes identically. The strong maximum principle is based on the Hopf boundary maximum principle; see Walter [1; p. 294]: B.5 Theorem. Let L be elliptic of type (B.1) on the ball B = BR (0) ⊂ IRn , with c ≥ 0. Suppose u ∈ C 2 (B) ∩ C 0 (B) satisﬁes Lu ≥ 0 in B, u ≥ 0 on ∂B, and u ≥ γ > 0 in Bρ (0) for some ρ < R, γ > 0. Then there exists δ = δ(L, γ, ρ, R) > 0 such that u(x) ≥ δ R − |x| in B . In particular, if u ∈ C 2 (B) ∩ C 1 (B) and if u(x0 ) = 0 for some x0 ∈ ∂B, then the interior normal derivative of u at the point x0 is strictly positive. Weak Maximum Principle For weak solutions of elliptic equations we have the following analogue of Theorem B.4. B.6 Theorem. Suppose L is elliptic of type (B.2) and suppose the Dirichlet form of L is positive deﬁnite on H01,2 (Ω) in the sense that L(u, u) > 0 for all u ∈ H01,2 (Ω), u = 0. Then, if u ∈ H 1,2 (Ω) weakly satisﬁes L u ≥ 0 in the sense that L(u, ϕ) ≥ 0 for all non-negative ϕ ∈ H01,2 (Ω),

Appendix B

273

and u ≥ 0 on ∂Ω, it follows that u ≥ 0 in Ω. Proof. Choose ϕ = u− = max {−u, 0} ∈ H01,2 (Ω). Then 0 ≤ L(u, u− ) = −L(u− , u− ) ≤ 0 with equality if and only if u− ≡ 0; that is u ≥ 0. Theorem B.6 can be used to strengthen the boundary maximum principle Theorem B.5: B.7 Theorem. Let L satisfy the hypotheses of Theorem B.6 in Ω = BR (0) = B ⊂ IRn with coeﬃcients aij ∈ C 1 . Suppose u ∈ C 2 (B) ∩ C 1 (B) satisﬁes Lu ≥ 0 in B, u ≥ 0 on ∂B and u ≥ γ > 0 in Bρ (0) forsome ρ < R, γ > 0. Then there exists δ = δ(L, γ, ρ, R) > 0 such that u(x) ≥ δ R − |x| in B. Proof. We proof of Walter [1; p. 294]. For large C > 0 the function adapt the v = exp C R2 − |x|2 − 1 satisﬁes Lv ≤ 0 in B \ Bρ (0). Moreover, for small ε > 0 the function w = εv satisﬁes w ≤ u for |x| ≤ ρ and |x| = R. Hence, Theorem B.6 – applied to u − w on B \ Bρ (0) – shows that u ≥ w in B \ Bρ (0).

Application As an application, consider the operator L = −Δ − δ, where δ < λ1 , the ¯ ﬁrst Dirichlet eigenvalue of −Δ on Ω. Let u ∈ H01,2 (Ω) or u ∈ C2 (Ω) ∩ C o (Ω) weakly satisfy L u ≤ Co in Ω, u ≤ 0 on ∂Ω, and choose v(x) = C C − |x − x0 |2 ¯ and L v ≥ C0 . with x0 ∈ Ω and C suﬃciently large to achieve that v > 0 on Ω Then w = v − u satisﬁes L w ≥ 0 in Ω, w > 0 on ∂Ω , and hence w is non-negative throughout Ω. Thus u≤v

in Ω .

More generally, results like Theorem B.4 or B.5 can be used to obtain L∞ - or even Lipschitz a-priori bounds of solution to elliptic boundary value problems by comparing with suitably constructed “barriers”.

Appendix C

In this appendix we discuss the issue of (partial) diﬀerentiability of variational integrals of the type (C.1) E(u) = F x, u(x), ∇u(x) dx, Ω

where u ∈ H (Ω), for simplicity. Diﬀerentiability properties will crucially depend on growth conditions for F . 1.2

Fr´echet-diﬀerentiability A functional E on a Banach space X is Fr´echet-diﬀerentiable at a point u ∈ X if there exists a bounded linear map DE(u) ∈ X ∗ , called the diﬀerential of E at u, such that |E(u + v) − E(u) − DE(u)v| →0 vX as vX → 0. E is of class C 1 , if the map u → DE(u) is continuous. C.1 Theorem. Let Ω ⊂⊂ IRn . Suppose F : Ω × IR × IRn → IR is measurable in ∂ F, Fp = x ∈ Ω, continuously diﬀerentiable in u ∈ IR and p ∈ IRn , with Fu = ∂u ∂ ∂p F , and the following growth conditions are satisﬁed: 2n , if n ≥ 3, (1◦) |F (x, u, p)| ≤ C(1 + |u|s1 + |p|2 ), where s1 ≤ n−2 ◦ s2 t2 (2 ) |Fu (x, u, p)| ≤ C(1 + |u| + |p| ), where t2 < 2, if n ≤ 2, respectively, n+2 where s2 ≤ n+2 n−2 , t2 ≤ n , if n ≥ 3, n ◦ (3 ) |Fp (x, u, p)| ≤ C(1 + |u|s3 + |p|), where s3 ≤ n−2 , if n ≥ 3. 1 1,2 Then (C.1) deﬁnes a C -functional E on H (Ω). Moreover, DE(u) is given by < v, DE(u) >= Fu (x, u, ∇u)v + Fp (x, u, ∇u) · ∇v dx . Ω

Theorem C.1 applies for example to the functional |u|p dx G(u) = Ω

with p ≤

2n n−2 ,

if n ≥ 3, or to Dirichlet’s integral 1 E(u) = |∇u|2 dx. 2 Ω

Appendix C

275

Theorem C.1 rests on a result by Krasnoselskii [1; Theorem I.2.1]. For simplicity, we state this result for g: Ω × IRm → IR. To ensure measurability functions of composed functions g x, u(x) , with u ∈ Lp , we assume g: Ω × IRm → IR is a Carath´eodory function; that is, g is measurable in x ∈ Ω and continuous in u ∈ IRm . C.2 Theorem. Suppose g: Ω × IRm → IR is a Carath´eodory function satisfying the growth condition (1◦) |g(x, u)| ≤ C(1 + |u|s ) for some s ≥ 1. Then the operator u → (g(·, u ·) is continuous from Lsp (Ω) into Lp (Ω) for any p, 1 ≤ p < ∞. Theorem C.2 asserts that Nemitskii operators – that is, evaluation operators like (C.1) – are continuous if they are bounded. For nonlinear operators this is quite remarkable. Using this result, Theorem C.1 follows quite naturally from the Sobolev embedding theorem, Theorem A.5. To get a ﬂavor of the proof, we establish continuity of the derivative of a functional E as in Theorem C.1. For u0 , u ∈ H01,2 (Ω), we estimate DE(u) − DE(u0 ) =

sup 1,2 v∈H 0 v 1,2 ≤1 H 0

| < v, DE(u) − DE(u0 ) > |

≤ sup

|Fu (x, u, ∇u) − Fu (x, u0 , ∇u0 )| |v| dx |Fp (x, u, ∇u) − Fp (x, u0 , ∇u0 )| |∇v| dx + sup v

Ω

v

Ω

2n

|Fu (x, u, ∇u) − Fu (x, u0 , ∇u0 )| n+2 dx

≤ sup v

n+2 2n

⎛ ⎞ n−2 2n 2n · ⎝ |v| n−2 dx⎠

Ω

Ω

⎛ ⎞ 12 + sup |Fp (x, u, ∇u) − Fp (x, u0 , ∇u0 )|2 dx · ⎝ |∇v|2 dx⎠ 1 2

v

Ω

Ω

if n ≥ 3 – which we will assume from now on for simplicity. Now, by Theorem A.5 the integrals involving v are uniformly bounded for v ∈ H01,2 (Ω) with vH 1,2 ≤ 1. By our growth conditions (2◦ ) and (3◦ ), moreover, Fu (respec0 tively Fp ) can be estimated like 2n 2n |Fu (x, u, ∇u)| n+2 ≤ C 1 + |u| n−2 + |∇u|2 , and by Theorem C.2 it follows that DE(u) → DE(u0 ), if u → u0 , as desired.

276

Appendix C

Natural Growth Conditions Conditions (1◦ )–(3◦ ) of Theorem C.1 require a special structure of the function F ; for instance, terms involving |∇u|2 cannot involve coeﬃcients depending on u. Consider for example the functional in Section I.1.5, given by F (x, u, p) = gij (u)pi pj . Note that Fu has the same growth as F with respect to p. More generally, for analytic functions F such that (C.2) |p|2 ≤ F (x, u, p) ≤ C |u| 1 + |p|2 one would expect the following growth conditions: (C.3) Fu (x, u, p) ≤ C |u| 1 + |p|2 (C.4) Fp (x, u, p) ≤ C |u| 1 + |p| for x ∈ Ω, u ∈ IR, and p ∈ IRn . Under these growth assumptions, in general a functional E given by (C.1) cannot be Fr´echet-diﬀerentiable in H 1,2 (Ω) any more. However, minimizers (in H01,2 (Ω), say) still may exist, compare Theorem I.1.5. Is it still possible to derive necessary conditions in the form of Euler-Lagrange equations? – The answer to this question may be positive, if we consider only a restricted set of minimizers and a narrower class of “testing functions”, that is of admissible variations: C.3 Theorem. Suppose E is given by (C.1) with a Carath´eodory function F , of class C 1 in u and p, satisfying the natural growth conditions (C.2)–(C.4). Then, if u, ϕ ∈ H 1,2 ∩ L∞ (Ω), the directional derivative of E at u in direction ϕ exists and is given by: d E(u + εϕ)ε=0 = Fu (x, u, ∇u)ϕ + Fp (x, u, ∇u) · ∇ϕ dx . dε Ω

In particular, at a minimizer u ∈ H 1,2 ∩ L∞ of E, with F satisfying (C.2)– (C.4), the Euler-Lagrange equations are weakly satisﬁed in the sense that Fu (x, u, ∇u) · ϕ + Fp (x, u, ∇u)∇ϕ dx = 0 Ω

holds for all ϕ ∈ H01,2 ∩ L∞ (Ω). Note that the assumption u ∈ L∞ often arises naturally, as in Theorem I.1.5. Sometimes, boundedness of minimizers may also be derived a posteriori. For further details, we refer to Giaquinta [1] or Morrey [4]. The question of differentiability of functionals in general is quite subtle, as is illustrated by an example of Ball and Mizel [1].

References

Acerbi, E.-Fusco, N. : [1] Semicontinuity problems in the calculus of variations. Arch. Rat. Mech. Anal. 86 (1984) 125–145 Adams, R.A. : [1] Sobolev spaces. Academic Press, New York-San Francisco-London (1975) Adimurthi : [1] Existence of positive solutions of the semilinear Dirichlet problem with critical growth for the n-Laplacian. Ann. Sc. Norm. Sup. Pisa, Ser. IV 17 (1990) 393–413 Adimurthi-Robert, F.-Struwe, M. : [1] Concentration phenomena for Liouville’s equation in dimension four. J. Eur. Math. Soc. (JEMS) 8 (2006), no. 2, 171–180 Adimurthi-Prashanth, S. : [1] Failure of Palais-Smale condition and blow-up analysis for the critical exponent problem in IR2 . Proc. Indian Acad. Sci. Math. Sci. 107 (1997), no. 3, 283–317 Adimurthi-Srikanth, P.N.-Yadava, S.L. : [1] Phenomena of critical exponent in IR2 . Proc. R. Soc. Edinburgh, Sect. A 119, No. 1/2 (1991) 19–25 Adimurthi-Struwe, M. : [1] Global compactness properties of semilinear elliptic equations with critical exponential growth. J. Funct. Anal. 175 (2000), no. 1, 125–167 Adimurthi-Yadava, S.L. : [1] Elementary proof of the nonexistence of nodal solutions for the semilinear elliptic equations with critical Sobolev exponent. Nonlin. Analysis, T.M.A. 14 (1990) 785– 787 Ahmad, S.-Lazer, A.C.-Paul, J.L. : [1] Elementary critical point theory and perturbations of elliptic boundary value problems at resonance. Indiana Univ. Math. J. 25 (1976) 933–944 Almgren, F.J. : [1] Plateau’s problem. Benjamin, New York-Amsterdam (1966) Almgren, F.-Simon, L. : [1] Existence of embedded solutions of Plateau’s problem. Ann Sc. Norm. Sup. Pisa, Cl. Sci. (4) 6 (1979) 447–495 Alt, H.W. : [1] Verzweigungspunkte von H-Fl¨ achen. Part I: Math. Z. 127 (1972) 333–362 Part II: Math. Ann. 201 (1973) 33–55 Amann, H. : [1] Lusternik-Schnirelman theory and nonlinear eigenvalue problems. Math. Ann. 199 (1972) 55–72 [2] On the number of solutions of nonlinear equations in ordered Banach spaces. J. Funct. Anal. 14 (1973) 346–384 [3] Existence and multiplicity theorems for semi-linear elliptic boundary value problems. Math. Z. 150 (1976) 281–295 [4] Saddle points and multiple solutions of diﬀerential equations. Math. Z. 169 (1979) 127–166 Amann, H.-Zehnder, E. : [1] Nontrivial solutions for a class of nonresonance problems and applications to nonlinear diﬀerential equations. Ann. Sc. Norm. Sup. Pisa Cl. Sci. (4) 7 (1980) 539–603

278

References

Ambrosetti, A. : [1] On the existence of multiple solutions for a class of nonlinear boundary value problems. Rend. Sem. Mat. Univ. Padova 49 (1973) 195–204 [2] Problemi variazionali in analisi non lineare. Boll. Unione Mat. Ital. (7) 2-A (1988) 169–188 [3] Branching points for a class of variational operators. J. d’Analyse Math. 76 (1998) 321–335 Ambrosetti, A.-Badiale, M. : [1] Homoclinics: Poincar´ e-Melnikov type results via a variational approach. Ann. Inst. Henri Poincar´ e Analyse Non Lin´ eaire 15 (1998) 233-252 [2] Variational perturbative methods and bifurcation of bound states from the essential spectrum. Proc. Roy. Soc. Edinburgh Sect. A 128 (1998), no. 6, 1131–1161 Ambrosetti, A.-Badiale, M.-Cingolani, S. : [1] Semiclassical states of nonlinear Schr¨ odinger equations. Arch. Rational Mech. Anal. 140 (1997), no. 3, 285–300 Ambrosetti, A.-Garcia Azorero, J.-Peral, I. : [1] Perturbation of Δu + u(N +2)/(N −2) = 0, the scalar curvature problem in IRN and related topics. J. Funct. Anal. 165 (1999), no. 1, 117–149 Ambrosetti, A.-Lupo, D. : [1] On a class of nonlinear Dirichlet problems with multiple solutions. Nonlin. Anal. Theory Meth. Appl. 8 (1984) 1145–1150 Ambrosetti, A.-Mancini, G. : [1] Sharp non-uniqueness results for some nonlinear problems. Nonlin. Anal. Theory Meth. Appl. 3 (1979) 635–645 [2] On a theorem by Ekeland and Lasry concerning the number of periodic Hamiltonian trajectories. J. Diﬀerential Equations 43 (1982), no. 2, 249–256 Ambrosetti, A.-Rabinowitz, P.H. : [1] Dual variational methods in critical point theory and applications. J. Funct. Anal. 14 (1973) 349–381 Ambrosetti, A.-Struwe, M. : ∗ [1] A note on the problem −Δu = λu + u|u|2 −2 . Manusc. math. 54 (1986) 373–379 [2] Existence of steady vortex rings in an ideal ﬂuid. Arch. Rat. Mech. Anal. 108 (1989) 97–109 Arioli, G. : [1] A note on a quasilinear elliptic eigenvalue problem. Electron. J. Diﬀerential Equations 47 (1999), 12 pp. (electronic) Arioli, G.-Gazzola, F. : [1] Existence and multiplicity results for quasilinear elliptic diﬀerential systems. Commun. Partial Diﬀerential Equations 25 (2000), no. 1-2, 125–153 Arz´ ela, C. : [1] Il principio di Dirichlet. Rend. Regia Accad. Bologna (1897) Atkinson, F.V.-Brezis, H.-Peletier, L.A. : [1] Solutions d’´ equations elliptiques avex exposant de Sobolev critique qui changent de signe. C.R. Acad. Sci. Paris 306 Ser.I (1988) 711–714 Aubin, Th. : [1] Probl` emes isop´ erim´ etriques et espace de Sobolev. J. Diﬀ. Geom. 11 (1976) 573–598 [2] Equations diﬀ´ erentielles nonlin´ eaires et probl` eme de Yamabe concernant la courbure scalaire. J. Math. Pures Appl. 55 (1976) 269–293 [3] Nonlinear analysis on manifolds. Monge-Amp`ere equations, Grundlehren 252, Springer, New York-Heidelberg-Berlin (1982) Bahouri, H.-G´ erard, P. : [1] High frequency approximation of solutions to critical nonlinear wave equations. Am. J. Math. 121 (1999), no. 1, 131–175

References

279

Bahri, A. : [1] Topological results on a certain class of functionals and applications. J. Funct. Anal. 41 (1981) 397–427 [2] Critical points at inﬁnity in some variational problems. Pitman Research Notes Math. 182, Longman House, Harlow (1989) Bahri, A.-Berestycki, H. : [1] A perturbation method in critical point theory and applications. Trans. Am. Math. Soc. 267 (1981) 1–32 [2] Forced vibrations of superquadratic Hamiltonian systems. Acta Math. 152 (1984) 143–197 Bahri, A.-Coron, J.-M. : [1] On a nonlinear elliptic equation involving the critical Sobolev exponent: the eﬀect of the topology of the domain. Commun. Pure Appl. Math. 41 (1988) 253–294 [2] The scalar-curvature problem on the standard three-dimensional sphere. J. Funct. Anal. 95 (1991) 106–172 Bahri, A.-Lions, P.-L. : [1] Morse-index of some min-max critical points. Commun. Pure Appl. Math. 41 (1988) 1027–1037 Ball, J.M. : [1] Convexity conditions and existence theorems in nonlinear elasticity. Arch. Rat. Mech. Anal. 63 (1977) 337–403 [2] Constitutive inequalities and existence theorems in nonlinear elastostatics. In: Nonlinear Analysis and Mechanics, Research Notes Math. 17, Pitman, London (1977) 187–241 Ball, J.M.-Murat, F. : [1] W 1,p -quasiconvexity and variational problems for multiple integrals. J. Funct. Anal. 58 (1984) 225–253 Bandle, C.-Flucher, M. : [1] Harmonic radius and concentration of energy, hyperbolic radius and Liouville’s equan+2

tions ΔU = eU and ΔU = U n−2 . SIAM Review 38-2 (1996) 191–238 Bangert, V. : [1] Geod¨ atische Linien auf Riemannschen Mannigfaltigkeiten. Jber. Dt. Math.-Verein 87 (1985) 39–66 Bartolo, P.-Benci, V.-Fortunato, D. : [1] Abstract critical point theorems and applications to some nonlinear problems with “strong” resonance at inﬁnity. Nonlin. Anal. Theory Meth. Appl. 7 (1983) 981–1012 Bartsch, T. : [1] Topological methods for variational problems with symmetries. Lect. Notes Math. 1560, Springer, Berlin-Heidelberg (1993) Bartz, J.-Struwe, M.-Ye, R. : [1] A new approach to the Ricci ﬂow on S 2 . Annali Sc. Norm. Sup. Ser. IV 21 (1994) 475–482 Bass, H.-Connell, E.H.-Wright, D. : [1] The Jacobian conjecture: Reduction of degree and formal expansion of the inverse. Bull. Am. Math. Soc. 7 (1982) 287–330 Benci, V. : [1] Some critical point theorems and applications. Commun. Pure Appl. Math. 33 (1980) 147–172 [2] A geometrical index for the group S 1 and some applications to the study of periodic solutions of ordinary diﬀerential equations. Commun. Pure Appl. Math. 34 (1981) 393–432 [3] On critical point theory for indeﬁnite functionals in the presence of symmetries. Trans. Am. Math. Soc. 274 (1982) 533–572 [4] A new approach to the Morse-Conley theory and some applications. Annali Mat. Pura Appl. (4) 158 (1991) 231–305

280

References

Benci, V.-Cerami, G.-Passaseo, D. : [1] On the number of positive solutions of some nonlinear elliptic problems, in Nonlinear Analysis – A tribute in honour of Giovanni Prodi. Quaderni Sc. Norm. Sup. Pisa (eds. A. Ambrosetti, A. Marino), Pisa (1991) 93–107 Benci, V.-Hofer, H.-Rabinowitz, P.H. : [1] A priori bounds for periodic solutions on hypersurfaces. In: Periodic solutions of Hamiltonian systems and related topics (eds. Rabinowitz, P.H. et al.), D. Reidel Publishing Co. (1987) Benci, V.-Rabinowitz, P.H. : [1] Critical point theorems for indeﬁnite functionals. Invent. Math. 52 (1979) 241–273 Bensoussan, A.-Boccardo, L.-Murat, F. : [1] On a nonlinear partial diﬀerential equation having natural growth terms and unbounded solution. Ann. Inst. H. Poincar´e 5 (1988) 347–364 Berestycki, H.-Lasry, J.M.-Mancini, G.-Ruf, B. : [1] Existence of multiple periodic orbits on star-shaped Hamiltonian surfaces. Commun. Pure Appl. Math. 38 (1985) 253–290 Berestycki, H.-Lions, P.-L. : [1] Nonlinear scalar ﬁeld equations. I. Existence of ground state, Arch. Rat. Mech. Anal. 82 (1983) 313–345 [2] Existence d’´ etats multiples dans des equations de champs scalaire non lin´eaires dans le cas de masse nulle. C. R. Acad. Sci. Paris 297 (1983), Ser. I 267–270 Berger, M.S. : [1] Nonlinearity and functional analysis. Academic Press, New York (1978) Berger, M.S.-Berger, M.S. : [1] Perspectives in nonlinearity. Benjamin, New York (1968) Berkowitz, L.D. : [1] Lower semi-continuity of integral functionals. Trans. Am. Math. Soc. 192 (1974) 51–57 Bertsch, M.-Dal Passo, R.-van der Hout, R. : [1] Nonuniqueness for the heat ﬂow of harmonic maps on the disk. Arch. Ration. Mech. Anal. 161 (2002) 93–112 Bethuel, F. : [1] The approximation problem for Sobolev maps between manifolds. Acta Math. 167 (1991) 153–206 Bethuel, F.-Brezis, H.-H´ elein, F. : [1] Ginzburg-Landau vortices. PNLDE 13, Birkh¨ auser, Boston-Basel-Berlin (1994) Birkhoﬀ, G.D. : [1] Dynamical systems with two degrees of freedom. Trans. Am. Math. Soc. 18 (1917) 199–300 [2] Dynamical systems. Am. Math. Soc. Coll. Publ. 9, Prividence (1927) Bobenko, A.-Kuksin, S. : [1] The nonlinear Klein-Gordon equation on an interval as a perturbed Sine-Gordon equation. Comm. Math. Helv. 70 (1995), no. 1, 63–112 Boccardo, L.-Murat, F.-Puel, J.P. : [1] Existence de solutions faibles pour des ´ equations elliptiques quasi-lin´ eaires a ` croissance quadratique. In: Nonlinear partial diﬀerential equations and their applications. College de France Seminar IV (ed.: Brezis-Lions), Research Notes Math. 84, Pitman, London (1983) 19–73 Bolle, P. : [1] On the Bolza problem. J. Diﬀ. Eq. 152 (1999) 274–288 Bolle, P.-Ghoussoub, N.-Tehrani, H. : [1] The multiplicity of solutions in non-homogeneous boundary value problems. Manuscripta Math. 101 (2000), no. 3, 325–350

References

281

B¨ ohme, R. : [1] Die L¨ osung der Verzweigungsgleichungen f¨ ur nichtlineare Eigenwertprobleme. Math. Z. 128 (1972) 105–126 Brendle, S. : [1] Convergence of the Yamabe ﬂow for arbitrary initial energy. J. Diﬀerential Geom. 69 (2005), no. 2, 217–278 Brezis, H. : [1] Operateurs maximaux monotone. Math. Studies, North-Holland, Amsterdam-London (1973) Brezis, H.-Coron, J.-M. : [1] Large solutions for harmonic maps in two dimensions. Commun. Math. Phys. 92 (1983) 203–215 [2] Multiple solutions of H-systems and Rellich’s conjecture. Commun. Pure Appl. Math. 37 (1984) 149–187 [3] Convergence of solutions of H-systems or how to blow bubbles. Arch. Rat. Mech. Anal. 89 (1985) 21–56 Brezis, H.-Coron, J.-M.-Nirenberg, L. : [1] Free vibrations for a nonlinear wave equation and a theorem of P. Rabinowitz. Commun. Pure Appl. Math. 33 (1980) 667–689 Brezis, H.-Kato, T. : [1] Remarks on the Schr¨ odinger operator with singular complex potentials. J. Math. Pures Appl. 58 (1979) 137–151 Brezis, H.-Merle, F. : [1] Uniform estimates and blow-up behavior for solutions of −Δu = V (x)eu in two dimensions. Commun. Partial Diﬀerential Equations 16 (1991) 1223–1253 Brezis, H.-Nirenberg, L. : [1] Forced vibrations for a nonlinear wave equation. Commun. Pure Appl. Math. 31 (1978) 1–30 [2] Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents. Commun. Pure Appl. Math. 36 (1983) 437–477 Brezis, H.-Waigner, S. : [1] A note on limiting cases of Sobolev imbeddings and convolution inequalities. Commun. Part. Diﬀ. Eq. 5 (1989) 773–789 Browder, F.E. : [1] Inﬁnite dimensional manifolds and nonlinear elliptic eigenvalue problems. Ann. Math. (2) 82 (1965) 459–477 [2] Existence theorems for nonlinear partial diﬀerential equations. Proc. Symp. Pure Math. 16, Am. Math. Soc. Prividence (1970) 1–60 [3] Nonlinear eigenvalue problems and group invariance. In: Functional analysis and related ﬁelds (ed.: Browder, F.E.), Springer (1970) 1–58 Campa, I.-Degiovanni, M. : [1] Subdiﬀerential calculus and nonsmooth critical point theory. Quaderno del Sem. Mat. di Brescia 15 1997 Canino, A.-Degiovanni, M. : [1] Nonsmooth critical point theory and quasilinear elliptic equations. Topological Methods in Diﬀerential Equations and Inclusions, 1-50 – A. Granas, M. Frigon, G. Sabidussi Eds. – Montreal 1994, NATO ASI Series – Kluwer A.P. 1995 Capozzi, A.-Fortunato, D.-Palmieri, G. : [1] An existence result for nonlinear elliptic problems involving critical Sobolev exponent. Ann. Inst. H. Poincar´ e Analyse Nonlin´ eaire 2 (1985) 463–470 Carleson, L.-Chang, S.Y.A : [1] On the existence of an extremal function for an inequality of J. Moser. Bull. Sci. Math. Astro. (2) 110 (1986) 113–127 Cerami, G. : [1] Un criterio di esistenza per i punti critici su variet` a illimitate. Rend. Acad. Sci. Let. Ist. Lombardo 112 (1978) 332–336

282 [2]

References Sull’ esistenza di autovalori per un problema al contorno non lineare. Ann. di Mat. (IV) 24 (1980) 161–179

Cerami, G.-Fortunato, D.-Struwe, M : [1] Bifurcation and multiplicity results for nonlinear elliptic problems involving critical Sobolev exponents. Ann. Inst. H. Poincar´ e, Analyse Nonlin´ eaire 1 (1984) 341–350 Cerami, G.-Solimini, S.-Struwe, M. : [1] Some existence results for superlinear elliptic boundary value problems involving critical exponents. J. Funct. Anal. 69 (1986) 289–306 Chang, K.C. : [1] Solutions of asymptotically linear operator equations via Morse theory. Commun. Pure Appl. Math. 34 (1981) 693–712 [2] Variational methods for non-diﬀerentiable functionals and their applications to partial diﬀerential equations J. Math. Anal. Appl. 80 (1981) 102–129 [3] Inﬁnite dimensional Morse theory and its applications. Lect. Notes 22nd Session Sem. Math. Sup. Montreal (1983) [4] A variant of the mountain pass lemma. Scientia Sinica 26 (1983) 1241–1255 [5] Variational methods and sub- and super-solutions. Sci. Sinica Ser. A 26 (1983) 1256– 1265 [6] Heat ﬂow and boundary value problem for harmonic maps. Ann. Inst. H. Poincar´e Analyse Non Lin´ eaire 6 (1989) 363–395 [7] Inﬁnite dimensional Morse theory and multiple solution problems. Progr. Nonlin. Diﬀ. Eq. Appl. 6, Birkh¨ auser, Boston (1993) Chang, K.C.-Ding, W.-Y. : [1] A result on the global existence for heat ﬂows of harmonic maps from D2 into S 2 . In: Nematics (Orsay, 1990), 37–47, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 332, Kluwer Acad. Publ., Dordrecht, 1991 Chang, K.C.-Ding, W.-Y.-Ye, R. : [1] Finite-time blow-up of the heat ﬂow of harmonic maps from surfaces. J. Diﬀ. Geom. 36 (1992) 507–515 Chang, K.C.-Eells, J. : [1] Unstable minimal surface coboundaries. Acta Math. Sinica (New Ser.) 2 (1986) 233– 247 Chang, S.-Y. A.-Yang, P. : [1] Prescribing Gaussian curvature on S 2 . Acta Math. 159 (1987) 215–259 Chen, W.-X.-Ding, W.-Y. : [1] Scalar curvatures on S 2 . Trans. Am. Math. Soc. 303 (1987) 365–382 Chen, Y.-Ding, W.-Y. : [1] Blow-up and global existence for heat ﬂows of harmonic maps. Invent. Math. 99 (1990) 567–578 Chen, Y.-Struwe, M. : [1] Existence and partial regularity results for the heat ﬂow for harmonic maps. Math. Z. 201 (1989) 83–103 Chow, S.-N.-Hale, J.K : [1] Methods of bifurcation theory. Grundlehren 251, Springer, New York-HeidelbergBerlin (1982) Cianchi, A. : [1] A sharp form of Poincar´ e-type inequalities on balls and spheres. ZAMP 40 (1989) 558-569 Ciarlet, P.G. : [1] Three-dimensional elasticity. Math. Elasticity 1, North-Holland, Amsterdam (1987) Clark, D.C. : [1] A variant of the Lusternik-Schnirelman theory. Indiana Univ. Math. J. 22 (1972) 65–74 Clarke, F.H. : [1] A new approach to Lagrange multipliers. Math. Oper. Res. 1 (1976) 165–174

References [2] [3] [4] [5]

283

Solution p´ eriodique des ´ equations hamiltoniennes. C. R. Acad. Sci. Paris 287 (1978) 951–952 A classical variational principle for Hamiltonian trajectories. Proc. Amer. Math. Soc. 76 (1979) 186–188 Periodic solutions to Hamiltonian inclusions. J. Diﬀ. Eq. 40 (1981) 1–6 Optimization and nonsmooth analysis. Wiley Interscience, New York (1983)

Clarke, F.H.-Ekeland, I. : [1] Hamiltonian trajectories having prescribed minimal period. Commun. Pure Appl. Math. 33 (1980) 103–116 Coﬀman, C.V. : [1] A minimum-maximum principle for a class of nonlinear integral equations. J. Analyse Math. 22 (1969) 391–419 Coifman, R.-Lions, P.-L.-Meyer, Y.-Semmes, S. : [1] Compensated compactness and Hardy spaces. J. Math. Pures Appl. 72 (1993) 247– 286 Conley, C. : [1] Isolated invariant sets and the Morse index. CBMS 38, Am. Math. Soc., Providence (1978) Conley, C.-Zehnder E. : [1] A Morse type index theory for ﬂows and periodic solutions to Hamiltonian systems. Commun. Pure Appl. Math. 37 (1984) 207–253 Connor, E.-Floyd, E.E. : [1] Fixed point free involutions and equivariant maps. Bull. Am. Math. Soc. 66 (1960) 416–441 Coron, J.M. : [1] Periodic solutions of a nonlinear wave equation without assumption of monotonicity. Math. Ann. 262 (1983) 273–285 [2] Topologie et cas limite des injections de Sobolev. C.R. Acad. Sc. Paris 299, Ser. I (1984) 209–212 [3] Nonuniqueness for the heat ﬂow of harmonic maps. Ann. Inst. M. Poincar´e Analyse Nonlin´ eaire 7 (1990) 335–344 Coron, J.M.-Ghidaglia, J.M. : [1] Explosion en temps ﬁni pour le ﬂot des applications harmoniques. C.R. Acad. Sci. Paris Ser. I Math. 308 (1989) 339–344 Corvellec, J.N.-Degiovanni, M.-Marzocchi, M. : [1] Deformation properties for continuous functionals and critical point theory. Top. Meth. Nonl. Anal. 1 (1993) 151–171 Coti Zelati, V.-Ekeland, I.-Sere, E. : [1] A variational approach to homoclinic orbits in Hamiltonian systems. Math. Ann. 288, No.1 (1990) 133–160 Coti Zelati, V.-Rabinowitz, P.H. : [1] Homoclinic type solutions for a semilinear elliptic PDE on IRn . Commun. Pure Appl. Math. 45, No.10 (1992) 1217–1269 Courant, R. : [1] Dirichlet’s principle, conformal mapping and minimal surfaces. Interscience, New York (1950). Reprinted: Springer, New York-Heidelberg-Berlin (1977) Crandall, M.G.-Rabinowitz, P.H. : [1] Continuation and variational methods for the existence of positive solutions of nonlinear elliptic eigenvalue problems. Arch. Rat. Mech. Anal. 58 (1975) 201–218 Dacorogna, B. : [1] Weak continuity and weak lower semicontinuity of non-linear functionals. Lect. Notes Math. 922, Springer, Berlin-Heidelberg-New York (1982) [2] Direct methods in the calculus of variations. Appl. Math. Sci. 78 (1989), Springer, Berlin-Heidelberg-New York (1989)

284

References

Dancer, E.N. : [1] Degenerate critical points, homotopy indices, and Morse inequalities. J. Reine Angew. Math. 350 (1984) 1–22 [2] Counterexamples to some conjectures on the number of solutions of nonlinear equations. Math. Ann. 272 (1985) 421–440 de Figueiredo, D.G. : [1] Lectures on the Ekeland variational principle with applications and detours. Lect. notes, College on Variational Problems in Analysis, Trieste (1988) Degiovanni, M.-Marzocchi, M. : [1] A critical point theory for nonsmooth functionals. Ann. Mat. Pura Appl. 4, 167 (1994) 73–100 Deimling, K. : [1] Nonlinear functional analysis. Springer, Berlin-Heidelberg-New York-Tokyo (1985) Del Pino, M.-Felmer, P.L. : [1] Local minimizers for the Ginzburg-Landau energy. Math. Z. 225 (1997), no. 4, 671– 684 Devillanova, G.-Solimini, S. : [1] Concentration estimates and multiple solutions to elliptic problems at critical growth. Adv. Diﬀerential Equations 7 (2002), no. 10, 1257–1280 Di Benedetto, E. : [1] C 1+α local regularity of weak solutions to degenerate elliptic equations. Nonlinear Analysis, T.M.A. 7 (1983) 827–850 Ding, W.-Y. : [1] On a conformally invariant elliptic equation on IRn . Commun. Math. Phys. 107 (1986) 331–335 Ding, W.-Y.-Jost, J.-Li, J.-Wang,G. : [1] Existence results for mean ﬁeld equations. Ann. Inst. H. Poincar Anal. Non Linaire 16 (1999), no. 5, 653–666 Ding, W.-Y.-Jost, J.-Li, J.-Peng, X.-Wang, G. : [1] Self duality equations for Ginzburg-Landau and Seiberg-Witten type functionals with 6th order potentials. Commun. Math. Phys. 217 (2001), no. 2, 383–407 Ding, W.-Y.-Tian, G. : [1] Energy identity for a class of approximate harmonic maps from surfaces. Commun. Anal. Geom. 3 (1995), no. 3-4, 543–554 Di Perna, R.J. : [1] Convergence of approximate solutions to conservation laws. Arch. Rat. Mech. Anal. 82 (1983) 27–70 Di Perna, R.J.-Majda, A.J. : [1] Oscillations and concentrations in weak solutions of the incompressible ﬂuid equations. Commun. Math. Phys. 108 (1987) 667–689 Dirichlet, L. : [1] Vorlesungen u ¨ber die im umgekehrten Verh¨ altnis des Quadrats der Entfernung wirkenden Kr¨ afte. G¨ ottingen (1856/57) Douglas, J. : [1] Solution of the problem of Plateau. Trans. Am. Math. Soc. 33 (1931) 263–321 Druet, O. : [1] Multibumps analysis in dimension 2: quantiﬁcation of blow-up levels. Duke Math. J. 132 (2006), no. 2, 217–269 Druet, O.-Hebey, E.-Robert, F. : [1] Blow-up theory for elliptic PDEs in Riemannian geometry. Mathematical Notes, 45. Princeton University Press, Princeton, NJ, 2004 Duc, D. M. : [1] Nonlinear singular elliptic equations. J. London Math. Soc. (2) 40 (1989) 420–440 [2] Flow methods in global analysis. Acta Math. Vietnam. 20, No.2 (1995) 171–180

References

285

Dunford, N.-Schwartz, J.T. : [1] Linear operators. Vol. I, Intescience, New York (1958) Edmunds, D.E.-Moscatelli, V.B. : [1] Sur la distribution asymptotique des valeurs propres pour une classe g´en´ erale d’op´ erateurs diﬀ´ erentiels. C.R. Acad. Sci. Paris 284 (1978) 1283–1285 Eells, J. : [1] A setting for global analysis. Bull. Am. Math. Soc. 72 (1966) 751–809 Eells, J.-Lemaire, L. : [1] Report on harmonic maps. Bull. London Math. Soc. 10 (1978) 1–68 [2] Another report on harmonic maps. Bull. London Math. Soc. 20 (1988) 385–524 Eells, J.-Sampson, J.H. : [1] Harmonic mappings of Riemannian manifolds. Am. J. Math. 86 (1964) 109–160 Eells, J.-Wood, J.C. : [1] Restrictions on harmonic maps of surfaces. Topology 15 (1976) 263–266 Egnell, H. : [1] Linear and nonlinear elliptic eigenvalue problems. Dissertation, Univ. Uppsala (1987) Eisen, G. : [1] A counterexample for some lower semicontinuity results. Math. Z. 162 (1978) 241–243 [2] A selection lemma for sequences of measurable sets, and lower semi-continuity of multiple integrals. Manusc. math. 27 (1979) 73–79 Ekeland, I. : [1] On the variational principle. J. Math. Anal. Appl. 47 (1974) 324–353 [2] Convexity methods in Hamiltonian mechanics. Ergebnisse d. Math. (Ser. III) 19, Springer (1990) Ekeland, I.-Ghoussoub, N. : [1] Selected new aspects of the calculus of variations in the large. Bull. Am. Math. Soc. (N.S.) 39 (2002), no. 2, 207–265 Ekeland, I.-Hofer, H. : [1] Periodic solutions with prescribed minimal period for convex autonomous Hamiltonian systems. Inv. Math. 81 (1985) 155–188 Ekeland, I.-Lasry, J.M. : [1] On the number of periodic trajectories for a Hamiltonian ﬂow on a convex energy surface. Ann. Math. 112 (1980) 283–319 Ekeland, I.-Lassoued, L. : [1] Un ﬂot hamiltonien a au moins deux trajectoires ferm´ees sur toute surface d’ ´ energie convexe et born´ ee. C. R. Acad. Sci. Paris 361 (1985) 161–164 Ekeland, I.-Temam, R. : [1] Convex analysis and variational problems. North Holland, Amsterdam (1976) Evans, L.C. : [1] Quasiconvexity and partial regularity in the calculus of variations. Arch. Rat. Mech. Anal. 95 (1986) 227–252 [2] Weak convergence methods for nonlinear partial diﬀerential equations. CBMS 74, Am. Math. Soc., Providence (1990) Evans, L.C.-Gariepy, R.F. : [1] Blowup, compactness and partial regularity in the calculus of variations. Indiana Univ. Math. J. 36 (1987) 361–371 Euler, L. : [1] Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes sive solutio problematis isoperimetrici latissimo sensu accepti. Lausanne-Gen` eve (1744), Opera, Ser. I, Vol. 24 (ed. C. Carath´eodory), Bern (1952) Fadell, E.R.-Husseini, S. : [1] Relative cohomological index theories. Advances Math. 64 (1987) 1–31

286

References

Fadell, E.R.-Husseini, S.-Rabinowitz, P.H. : [1] Borsuk-Ulam theorems for arbitrary S 1 actions and applications. Trans. Am. Math. Soc. 274 (1982) 345–360 Fadell, E.R.-Rabinowitz, P.H. : [1] Bifurcation for odd potential operators and an alternative topological index. J. Funct. Anal. 26 (1977) 48–67 [2] Generalized cohomological index theories for Lie group actions with applications to bifurcation questions for Hamiltonian systems. Invent. Math. 45 (1978) 139–174 Federer, H. : [1] Geometric measure theory. Grundlehren 153, Springer, Berlin-Heidelberg-New York (1969) Feﬀerman, C.-Stein, E.M. : [1] H p -spaces of several variables. Acta Math. 129 (1972) 137–193 Fet, A.I. : [1] Variational problems on closed manifolds. Mat. Sb. (N.S.) 30 (1952) 271–316 (Russian) Am. Math. Soc. Transl. 90 (1953) Floer, A. : [1] Morse theory for ﬁxed points of symplectic diﬀeomorphisms. Bull. Am. Math. Soc. 16 (1987) 279–281 [2] Symplectic ﬁxed points and holomorphic spheres. Commun. Math. Phys. 120 (1989) 575–611 Flucher, M. : [1] Extremal functions for the Trudinger-Moser inequality in two dimensions. Comm. Math. Helv. 67 (1992) 471–497 [2] Variational problems with concentration and the role of Robin functions. Habilitationsschrift 1996 Fortunato, D.-Giannoni, F.-Masiello, A. : [1] A Fermat principle for stationary space-times and applications to light rays. J. Geometry Phys. 515 (1994) 1–30 Fortunato, D.-Jannelli, E. : [1] Inﬁnitely many solutions for some nonlinear elliptic problems in symmetrical domains. Proc. Royal Soc. Edinburgh 105A (1987) 205–213 Fraenkel, L.E.-Berger, M.S. : [1] A global theory of steady vortex rings in an ideal ﬂuid. Acta Math. 132 (1974) 13–51 Franks, J. : [1] Geodesics on S 2 and periodic points of annulus homeomorphisms. Invent. Math. 108 (1992) 403–418 Frehse, J. : [1] Capacity methods in the theory of partial diﬀerential equations. Jber. d. Dt. Math.Verein. 84 (1982) 1–44 [2] Existence and perturbation theorems for non-linear elliptic systems. In: Nonlinear partial diﬀerential equations and their applications, College de France Seminar IV (ed.: Brezis-Lions), Research Notes Math. 84, Pitman, London (1983) Freire, A. : [1] Uniqueness for the harmonic map ﬂow in two dimensions. Cal. Var. 3 (1994) 95–105 [2] Uniqueness for the harmonic map ﬂow from surfaces to general targets. Comm. Math. Helv. 70 (1995) 310–338; Correction. Comment. Math. Helv. 71, No.2 (1996) 330–337 Fusco, N. : [1] Quasi convessit` a e semicontinuit` a per integrali multipli di ordine superiore. Ricerche di Mat. 29 (1980) 307–323 Gauss, C.F. : [1] Allgemeine Lehrs¨ atze in Beziehung auf die im verkehrten Verh¨ altnisse des Quadrats der Entfernung wirkenden Anziehungs- und Abstossungskr¨ afte. Leipzig (1840); Werke V, G¨ ottingen (1867) 195–242

References

287

Gehring, F.W. : [1] The Lp -integrability of the partial derivatives of a quasi-conformal mapping. Acta Math. 130 (1973) 265–277 Gelfand, I.M. : [1] Some problems in the theory of quasilinear equations. Am. Math. Soc. Transl., Ser. 2, 29 (1963) 295–381 Ghoussoub, N. : [1] Location, multiplicity and Morse indices of min-max critical points. J. Reine Angew. Math. 417 (1991) 27–76 [2] Duality and perturbation methods in critical point theory. Cambridge Tracts in Math. 107, Cambridge University Press, Cambridge, (1993) Giannoni, F. : [1] Multiple brake orbits in a potential well and a Seifert conjecture. Proceedings of the Third World Congress of Nonlinear Analysts, Part 5 (Catania, 2000). Nonlinear Anal. 47 (2001), no. 5, 3073–3084 Giannoni, F.-Masiello, A.-Piccione, P. : [1] A timelike extension of Fermat’s principle in general relativity and applications. Calc. Var. 6 (1998) 263–282 Giaquinta, M. : [1] Multiple integrals in the calculus of variations and nonlinear elliptic systems. Ann. Math. Studies 105, Princeton Univ. Press, Princeton (1983) Giaquinta, M.-Giusti, E. : [1] On the regularity of the minima of variational integrals. Acta Math. 148 (1982) 31–46 Giaquinta, M.-Modica, G. : [1] Regularity results for some classes of higher order nonlinear elliptic systems. J. reine angew. Math. 311/312 (1979) 145–169 [2] Partial regularity of minimizers of quasiconvex integrals. Ann. Inst. H. Poincar´e, Anal. Non Lin´ eaire 3 (1986) 185–208 Giaquinta, M.-Modica, G.-Souˇ cek, J. : [1] Cartesian currents, weak diﬀeomorphism and existence theorems in nonlinear elasticity. Arch. Rat. Mech. Anal. 106 (1989) 97–159, erratum and addendum: Arch. Rat. Mech. Anal. 109 (1990) 385–392 Gidas, B.-Ni, W.-M.-Nirenberg, L. : [1] Symmetry and related properties via the maximum principle. Commun. Math. Phys. 68 (1979) 209–243 Gilbarg, D.-Trudinger, N.S. : [1] Elliptic partial diﬀerential equations of second order. 2nd edition, Grundlehren 224, Springer, Berlin-Heidelberg-New York-Tokyo (1983) Ginzburg, V.L. : [1] An embedding S 2n−1 → IR2n , 2n − 1 ≥ 7, whose Hamiltonian ﬂow has no periodic trajectories. IMRN 2 (1995), 83–98 Giusti, E. : [1] Minimal surfaces and functions of bounded variation. Monographs in Mathematics 80, Birkh¨ auser, Boston-Basel-Stuttgart (1984) Girardi, M.-Matzeu, M. : [1] Solutions of minimal period for a class of nonconvex Hamiltonian systems and applications to the ﬁxed energy problem. Nonlinear Analysis T.M.A. 10 (1986) 371–382 [2] Periodic solutions of convex autonomous Hamiltonian systems with a quadratic growth at the origin and superquadratic at inﬁnity. Ann. di Mat. Pura Appl. (IV) 147 (1987) 21–72 Goldstine, H.H. : [1] A history of the calculus of variations from the 17th through the 19th century. Studies Hist. Math. Phys. Sci. 5, Springer, New York-Heidelberg-Berlin (1980) Gr¨ uneberg, M. : [1] The Yamabe ﬂow on three-manifolds. Ph.D. Thesis, Stanford (2004)

288

References

Guedda, M.-Veron, L. : [1] Quasilinear elliptic equations involving critical Sobolev exponents. Nonlinear Analysis, T.M.A. 13 (1989) 879–902 Gulliver, R.D. : [1] Regularity of minimizing surfaces of prescribed mean curvature. Ann. Math. (2) 97 (1973) 275–305 Gulliver, R.D.-Lesley, F.D. : [1] On boundary branch points of minimizing surfaces. Arch. Rat. Mech. Anal. 52 (1973) 20–25 Gulliver, R.D.-Osserman, R.-Royden, H.L. : [1] A theory of branched immersions of surfaces. Am. J. Math. 95 (1973) 750–812 G¨ unther, M. : [1] On the perturbation problem associated to isometric embeddings of Riemannian manifolds. Ann. Global Anal. Geom. 7 (1989) 69–77 [2] Conformal normal coordinates. Ann. Global Anal. Geom. 11 (1993), no. 2, 173–184 Hamilton, R. : [1] Harmonic maps of manifolds with boundary. Lect. Notes Math 471, Springer, Berlin (1975) [2] Lectures on geometric ﬂows. (1989), unpublished Hang, Fengbo-Lin, Fanghua : [1] Topology of Sobolev mappings. Math. Res. Lett. 8 (2001), no. 3, 321–330 Harpes, P. : [1] Ginzburg-Landau type approximations for the Landau-Lifshitz ﬂow and the harmonic map ﬂow in two space dimensions. Ph.D. Thesis, ETH-Z¨ urich (2001) Hartman, P.-Wintner, A. : [1] On the local behavior of solutions of non-parabolic partial diﬀerential equations. Am. J. Math. 75 (1953) 449–476 Hebey, E.-Robert, F. : [1] Coercivity and Struwe’s compactness for Paneitz type operators with constant coefﬁcients. Calc. Var. Partial Diﬀerential Equations 13 (2001), no. 4, 491–517 Heinz, E. : ¨ [1] Uber die Existenz einer Fl¨ ache konstanter mittlerer Kr¨ ummung bei vorgegebener Berandung. Math. Ann. 127 (1954) 258–287 ¨ [2] Uber die Eindeutigkeit beim Cauchyschen Anfangswertproblem einer elliptischen Differentialgleichung zweiter Ordnung. Nachr. Akad. Wiss. G¨ ottingen 1 (1955) 1–12 H´ elein, F. : [1] Regularit´ e des applications faiblement harmoniques entre une surface et une variet´e Riemannienne. C.R. Acad. Sci. Paris Ser. I Math. 312 (1991) 591–596 Hempel, J.A. : [1] Multiple solutions for a class of nonlinear elliptic boundary value problems. Indiana Univ. Math. J. 20 (1971) 983–996 Herman, M. : [1] Examples of compact hypersurfaces in IR2p , 2p ≥ 6, with no periodic orbits. fax to H. Hofer, Dec. 7, 1994 Hilbert, D. : ¨ [1] Uber das Dirichletsche Prinzip. Jber. Deutsch. Math. Vereinigung 8 (1900) 184–188 ¨ [2] Uber das Dirichletsche Prinzip. Festschrift zur Feier des 150-j¨ ahrigen Bestehens der K¨ oniglichen Gesellschaft der Wissenschaften zu G¨ ottingen 1901, Math. Ann. 49 (1904) 161–186 Hildebrandt, S. : [1] Boundary behavior of minimal surfaces. Arch. Rat. Mech. Anal. 35 (1969) 47–82 [2] On the Plateau problem for surfaces of constant mean curvature. Commun. Pure Appl. Math. 23 (1970) 97–114 [3] Nonlinear elliptic systems and harmonic mappings. Proc. Beijing Symp. Diﬀ. Geom. and Diﬀ. Eqs. (1980), Gordon and Breach (1983) 481–615

References [4]

289

The calculus of variations today, as reﬂected in the Oberwolfach meetings. In Perspect. Math., Anniversary of Oberwolfach 1984, Birkh¨ auser (1985)

Hildebrandt, S.-Tromba, A. : [1] Mathematics and optimal form. Scientiﬁc American Books, Inc. New York (1985) Hildebrandt, S.-Widman, K.-O. : [1] On the H¨ older continuity of weak solutions of quasilinear elliptic systems of second order. Ann. Sc. Norm. Sup. Pisa, Ser. 4, 4 (1977) 146–178 Hofer, H. : [1] Variational and topological methods in partially ordered Hilbert spaces. Math. Ann. 261 (1982) 493–514 [2] A geometric description of the neighborhood of a critical point given by the mountain pass theorem. J. London Math. Soc. (2) 31 (1985) 566–570 Hofer, H.-Zehnder, E. : [1] Periodic solutions on hypersurfaces and a result by C. Viterbo. Invent. Math. 90 (1987) 1–9 [2] Symplectic invariants and Hamiltonian dynamics. Birkh¨ auser Advanced Texts, Birkh¨ auser, Basel-Boston-Berlin (1994) Ioﬀe, A. : [1] Approximate subdiﬀerentials and applications. The metric theory. Mathematika 36 (1989) 1–38 Ioﬀe, A.-Schwartzman, E. : [1] Metric critical point theory 1. Morse regularity and homotopic stability of a minimum. J. Math. Pures Appl. 75 (1996) 125–153 Jeanjean, L. : [1] On the existence of bounded Palais-Smale sequences and application to a LandesmanLazer type problem set on IRN . Proc. Roy. Soc. Edinburgh Sect. A 129 (1999), no. 4, 787–809 Jeanjean, L.-Toland, J.F. : [1] Bounded Palais-Smale mountain-pass sequences. C.R. Acad. Sci. Paris, t. 327, S´ erie I (1998) 23–28 Jost, J. : [1] The Dirichlet problem for harmonic maps from a surface with boundary onto a 2sphere with non-constant boundary values. J. Diﬀ. Geom. 19 (1984) 393–401 [2] Harmonic maps between surfaces. Lect. Notes Math. 1062, Springer, Berlin-Heidelberg-New York (1984) [3] Nonlinear methods in Riemannian and K¨ ahlerian geometry. Second edition. DMV Seminar 10, Birkh¨ auser Verlag, Basel (1991) Jost, J.-Struwe, M. : [1] Morse-Conley theory for minimal surfaces of varying topological type. Invent. Math. 102 (1990) 465–499 Kazdan, J.-Warner, F. : [1] Scalar curvature and conformal deformation of Riemannian structure. J. Diﬀ. Geom. 10 (1975) 113–134 [2] Existence and conformal deformations of metrics with prescribed Gaussian and scalar curvatures. Ann. Math. 101 (1975) 317–331 Kelley, J.L. : [1] General topology. Graduate Texts Math. 27, Springer, New York-Heidelberg-Berlin (1955) Kinderlehrer, D.-Stampacchia, G. : [1] An introduction to variational inequalities and their applications. Academic Press, New York-London (1980) Klingenberg, W. : [1] Lectures on closed geodesics. Grundlehren 230, Springer, Berlin-Heidelberg-New York (1978)

290

References

Krasnoselskii, M.A. : [1] Topological methods in the theory of nonlinear integral equations. Macmillan, New York (1964) Krasnoselskii, M.A.-Zabreiko, P.P. : [1] Geometrical methods in nonlinear analysis. Grundlehren 263, Springer, Berlin-Heidelberg-New York-Tokyo (1984) Kuˇ cera, M. : [1] A new method for obtaining eigenvalues of variational inequalities based on bifurcaˇ tion theory. Cas. Pˇ est. Mat. 104 (1979) 389–411 Kuperberg, K. : [1] A smooth counterexample to the Seifert conjecture in dimension three. Ann. Math. 140 (1994) 723–732 Ladyzhenskaya, O.A. : [1] The mathematical theory of viscous imcompressible ﬂow. 2nd edition, Gordon & Breach, New York-London-Paris (1969) Ladyzhenskaya, O.A.-Solonnikov, V.A.-Ural’ceva, N.N. : [1] Linear and quasilinear equations of parabolic type. Am. Math. Soc. Transl. Math. Monogr. 23, Providence (1968) Ladyzhenskaya, O.A.-Ural’ceva, N.N. : [1] Linear and quasilinear elliptic equations. Academic Press, New York (1968) Lazer, A.-Solimini, S. : [1] Nontrivial solutions of operator equations and Morse indices of critical points of minmax type. Nonlinear Analysis, T.M.A. 12 (1988) 761–775 Lebesgue, H. : [1] Sur le probl` eme de Dirichlet. Rend. Circ. Mat. Palermo 24 (1907) 371–402 Lee, J.M.-Parker, T.H. : [1] The Yamabe problem. Bull. Am. Math. Soc. 17 (1987) 37–91 Lemaire, L. : [1] Applications harmoniques de surfaces riemanniennes. J. Diﬀ. Geom. 13 (1978) 51–78 L´ evy, P. : [1] Th´ eorie de l’addition des variables al´ eatoirs. Gauthier-Villars, Paris (1954) Li, Y.-Shafrir, I : [1] Blow-up analysis for solutions of −Δu = V eu in dimension two. Indiana Univ. Math. J. 43, No. 4 (1994) 1255–1270 Lin, F.-Wang, C. : [1] Energy identity of harmonic map ﬂows from surfaces at ﬁnite singular time. Calc. Var. Partial Diﬀerential Equations 6 (1998), no. 4, 369–380 Lions, J.L.-Magenes, E. : [1] Non-homogeneous boundary value problems and applications I. Grundlehren 181, Springer, Berlin-Heidelberg-New York (1972) Lions, P.-L. : [1] The concentration-compactness principle in the calculus of variations. The locally compact case. Part 1, Ann. Inst. H. Poincar´ e 1 (1984) 109–145 [2] The concentration-compactness principle in the calculus of variations. The locally compact case. Part 2, Ann. Inst. H. Poincar´ e 1 (1984) 223–283 [3] The concentration-compactness principle in the calculus of variations. The limit case. Part 1, Rev. Mat. Iberoamericano 1.1 (1985) 145–201 [4] The concentration-compactness principle in the calculus of variations. The limit case. Part 2, Rev. Mat. Iberoamericano 1.2 (1985) 45–121 Lions, P.-L.-Pacella, F.-Tricarico, M. : [1] Best constants in Sobolev inequalities for functions vanishing on some part of the boundary and related questions. Indiana Univ. Math. J. 37 (1988) 301–324 Loewner, C.-Nirenberg. L. : [1] Partial diﬀerential equations invariant under conformal or projective transformations, in: Contributions to Analysis, Academic Press (1974) 245–272

References

291

Lojasiewicz, S. : [1] Ensembles semi-analytiques. I.H.E.S. notes (1965) [2] Sur les ensembles semi-analytiques. Actes du Congrs International des Mathmaticiens (Nice, 1970), Tome 2, pp. 237–241. Gauthier-Villars, Paris, 1971 Long, Y. : [1] Multiple solutions of perturbed superquadratic second order Hamiltonian systems. Trans. Am. Math. Soc. 311 (1989) 749–780 Lovicarov´ a, H. : [1] Periodic solutions for a weakly nonlinear wave equation in one dimension. Czech. Math. J. 19 (1969) 324–342 Lusternik, L.-Schnirelman, L. : [1] Sur le probl` eme de trois g´ eod´ esiques ferm´ ees sur les surfaces de genre 0. C.R. Acad. Sci. Paris 189 (1929) 269–271 [2] M´ ethodes topologiques dans les probl`emes variationelles. Actualites Sci. Industr. 188, Paris (1934) Lusternik, L. : [1] The topology of function spaces and the calculus of variations in the large. Trudy Mat. Inst. Steklov 19 (1947) (Russian); Transl. Math. Monographs 16, Am. Math. Soc., Prividence (1966) Mancini, M.-Musina, R. : [1] A free boundary problem involving limiting Sobolev exponents. Manusc. math. 58 (1987) 77–93 Marcellini, P.-Sbordone, C. : [1] On the existence of minima of multiple integrals of the calculus of variations. J. Math Pures Appl. 62 (1983) 1–9 Marino, A. : [1] La biforcazione nel caso variazionale. Confer. Sem. Mat. Univ. Bari 132 (1973) Maupertuis, P.L.M. de : [1] Accord de diﬀ´ erentes lois de la nature qui avaient jusqu’ici paru incompatibles. M´em. Acad. Sci. (1744) 417–426 Mawhin, J. : [1] Probl` emes de Dirichlet variationels non lin´ eaires. Seminaire Math. Sup. 104, Presses Univ. Montreal, Montreal (1987) Mawhin, J.-Willem, M. : [1] Critical point theory and Hamiltonian systems. Appl. Math. Sci. 74, Springer, New York-Berlin-Heidelberg-London-Paris-Tokyo (1989) Mazzeo, R.-Pacard, F. : [1] A construction of singular solutions for a semilinear elliptic equation using asymptotic analysis. J. Diﬀer. Geom. 44, No.2 (1996) 331–370 Mazzeo, R.-Smale, N. : [1] Conformally ﬂat metrics of constant positive scalar curvature on subdomains of the sphere. J. Diﬀ. Geom. 34 (1991) 581–621 Meeks, W.H.-Yau, S.-T. : [1] The classical Plateau problem and the topology of three-dimensional manifolds. The embedding of the solution given by Douglas-Morrey and an analytic proof of Dehn’s lemma. Topology 21 (1982), no. 4, 409–442 Miersemann, E. : ¨ [1] Uber h¨ ohere Verzweigungspunkte nichtlinearer Variationsungleichungen. Math. Nachrichten 85 (1978) 195–213 Milnor, J. : [1] Morse theory. Ann. Math. Studies 51, Princeton Univ. Press, Princeton (1963) Morgan, F. : [1] Geometric measure theory. A beginner’s guide. Academic Press, London (1988)

292

References

Morrey, C.B.Jr. : [1] Existence and diﬀerentiability theorems for the solutions of variational problems for multiple integrals. Bull. Am. Math. Soc. 46 (1940) 439–458 [2] The problem of Plateau on a Riemannian manifold. Ann. Math. 49 (1948) 807–851 [3] Quasiconvexity and the lower semicontinuity of multiple integrals. Paciﬁc J. Math. 2 (1952) 25–53 [4] Multiple integrals in the calculus of variations. Springer Grundlehren 130, New York (1966) Morse, M. : [1] The critical points of functions and the calculus of variations in the large. Bull. Am. Math. Soc. 35 (1929) 38–54 [2] The calculus of variations in the large. Am. Math. Soc. Coll. Publ. 18 (1934) [3] Functional topology and abstract variational theory. Ann. Math. 38 (1937) 386–449 Morse, M.-Tompkins, C.B. : [1] The existence of minimal surfaces of general critical types. Ann. Math. 40 (1939) 443–472 [2] Unstable minimal surfaces of higher topological structure. Duke Math. J. 8 (1941) 350–375 Moser, J. : [1] A new proof of De Giorgi’s theorem. Commun. Pure Appl. Math. 13 (1960) 457–468 [2] A Harnack inequality for parabolic diﬀerential equations. Commun. Pure Appl. Math. 17 (1964) 101–134 [3] A sharp form of an inequality by N. Trudinger. Indiana Univ. Math. J. 20 (1971) 1077–1092 [4] Stable and random motions in dynamical systems. Ann. Math. Studies 77, Princeton Univ. Press, Princeton (1973) [5] On a nonlinear problem in diﬀerential geometry; in: Dynamical Systems (ed. M. Peixoto), Academic Press, New York (1973) M¨ uller, S. : [1] Weak continuity of determinants and nonlinear elasticity. C.R. Acad. Sci. Paris 307, s´ er. 1 (1988) 501–506 [2] A surprising higher integrability property of mappings with positive determinant. Bull. Am. Math. Soc. (N.S.) 21 (1989), no. 2, 245–248 [3] Higher integrability of determinants and weak convergence in L1 . J. Reine Angew. Math. 412 (1990), 20–34 Murat, F. : [1] Compacit´ e par compensation. Ann. Sc. Norm. Sup. Pisa (IV) 5 (1978) 489–507 [2] Compacit´ e par compensation II. in: Proc. intern. meet. recent methods in non-linear analysis, Rome, Pitagora ed. (1978) 245–256 Nash, J. : [1] The imbedding problem for Riemannian manifolds. Ann. Math. 63 (1956) 20–63 Neˇ cas, J. : [1] Les m´ ethodes directes en th´ eorie des ´ equations elliptiques. Academia, Prague (1967) Nehari, Z. : [1] Characteristic values associated with a class of nonlinear second-order diﬀerential equations. Acta Math. 105 (1961) 141–175 Ni, W.-M. : [1] Some minimax principles and their applications in nonlinear elliptic equations. J. Analyse Math. 37 (1980) 248–275 Nirenberg, L. : [1] Variational and topological methods in nonlinear problems. Bull. Am. Math. Soc. (New Ser.) 4 (1981) 267–302 Nitsche, J.C.C. : [1] The boundary behavior of minimal surfaces, Kellog’s theorem and branch points on the boundary. Invent. Math. 8 (1969) 313–333 Addendum, Invent. Math. 9 (1970) 270

References [2]

293

Vorlesungen u ¨ber Minimalﬂ¨ achen. Grundlehren 199, Springer, Berlin-HeidelbergNew York (1975)

Obata, M. : [1] Certain conditions for a Riemannian manifold to be isometric with a sphere. J. Math. Soc. Japan 14 (1962) 333–340 Ogawa, T.-Suzuki, T. : [1] Two dimensional elliptic equation with critical nonlinear growth. Trans. Am. Math. Soc. 350, no. 12 (1998) 4897–4918 Osserman, R. : [1] A survey of minimal surfaces. Math. Studies 25, Van Nostrand Reinhold Co., New York (1969) [2] A proof of the regularity everywhere of the classical solution to Plateau’s problem. Ann. Math. 91 (1970) 550–569 Palais, [1] [2] [3] [4]

R.S. : Morse theory on Hilbert manifolds. Topology 2 (1963) 299–340 Lusternik-Schnirelman theory on Banach manifolds. Topology 5 (1966) 115–132 Foundations of global non-linear analysis. Benjamin, New York (1968) Critical point theory and the minimax principle. Proc. Symp. Pure Math. 15 (1970) 185–212

Palais, R.S.-Smale, S. : [1] A generalized Morse theory. Bull. Am. Math. Soc. 70 (1964) 165–171 Parker, T.H. : [1] Bubble tree convergence for harmonic maps. J. Diﬀ. Geom. 44 (1996) 595–633 Patnaik, U. : [1] Volume constrained Douglas problem and the stability of liquid bridges between two coaxial tubes. Ph. D. Thesis, Univ. of Toledo (1994) Pohoˇ zaev, S. : [1] Eigenfunctions of the equation Δu + λf (u) = 0. Soviet Math. Dokl. 6 (1965) 1408– 1411 Poincar´ e, H. : [1] Les m´ ethodes nouvelles de la m´ echanique c´ eleste. Tome I, Dover Publications, New York (1957); Engl. transl.: National Aeronautics Space Administration, Springﬁeld (1967) Polya, G.-Szeg¨ o, G. : [1] Isoperimetric inequalities in mathematical physics. Ann. Math. Studies 27, Princeton Univ. Press, Princeton (1951) P¨ oschel, J. : [1] Quasi-periodic solutions for a nonlinear wave equation. Comm. Math. Helv. 71, No.2 (1996) 269–296 Pucci, P.-Serrin, J. : [1] Extensions of the mountain pass theorem. J. Funct. Anal. 59 (1984) 185–210 [2] The structure of the critical set in the mountain pass theorem. Trans. Am. Math. Soc. 299 (1987) 115–132 Pugh, C.-Robinson, C. : [1] The C 1 -closing lemma, including Hamiltonians. Ergodic Theory Dyn. Syst. 3 (1983) 261–313 Qing, J. : [1] Boundary regularity of weakly harmonic maps from surfaces. J. Funct. Anal. 114 (1993) 458–466 [2] On singularities of the heat ﬂow for harmonic maps from surfaces into spheres. Commun. Anal. Geom. 3 (1995) 297–315 Qing, J.-Tian, G : [1] Bubbling of the heat ﬂows for harmonic maps from surfaces. Commun. Pure Appl. Math. 50 (1997) 295–310

294

References

Quittner, P. : [1] Solvability and multiplicity results for variational inequalities. Comm. Math. Univ. Carolina 30 (1989) 281–302 Rabinowitz, P.H. : [1] Some aspects of nonlinear eigenvalue problems. Rocky Mount. J. Math. 3 (1973) 161–202 [2] Variational methods for nonlinear elliptic eigenvalue problems. Indiana Univ. Math. J. 23 (1974) 729–754 [3] Variational methods for nonlinear eigenvalue problems. In: Eigenvalues of nonlinear problems (ed. Prodi, G.), C.I.M.E., Edizioni Cremonese, Roma (1975) 141–195 [4] A variational method for ﬁnding periodic solutions of diﬀerential equations. Proc. Symp. Nonlinear Evol. Eq. (ed. Crandall, M.G.), Wisconsin (1977) 225–251 [5] Periodic solutions of Hamiltonian systems. Commun. Pure Appl. Math. 31 (1978) 157–184 [6] Free vibrations for a semilinear wave equation. Commun. Pure appl. Math. 31 (1978) 31–68 [7] Some minimax theorems and applications to nonlinear partial diﬀerential equations. In : Nonlinear analysis; a collection of papers in honor of Erich Rothe (ed. Cesari, L. et al.), Academic Press, New York (1978) 161–177 [8] Some critical point theorems and applications to semilinear elliptic partial diﬀerential equations. Ann. Sc. Norm. Sup. Pisa, Ser. 4, 5 (1978) 215–223 [9] Periodic solutions of Hamiltonian systems: A survey. Siam J. Math. Anal. 13 (1982) 343–352 [10] Multiple critical points of perturbed symmetric functionals. Trans. Am. Math. Soc. 272 (1982) 753–770 [11] Minimax methods in critical point theory with applications to diﬀerential equations. CBMS Regional Conference Series Math. 65, Am. Math. Soc., Providence (1986) [12] Homoclinic and heteroclinic orbits for a class of Hamiltonian systems. Calc. Var. Partial Diﬀer. Equ. 1, No.1 (1993) 1–36 [13] Solutions of heteroclinic type for some classes of semilinear elliptic partial diﬀerential equations. J. Math. Sci., Tokyo 1, No.3 (1994) 525–550 Rad´ o, T. : [1] On Plateau’s problem. Ann. Math. 31 (1930) 457–469 [2] The isoperimetric inequality and the Lebesgue deﬁnition of surface area. Trans. Am. Math. Soc. 61 (1947) 530–555 Rentzmann, S. : [1] Fl¨ achen konstanter mittlerer Kr¨ ummung und der Eﬀekt der Topologie. Diploma Thesis, ETH Z¨ urich, 1994 Reshetnyak, Y. : [1] General theorems on semicontinuity and on convergence with a functional. Sibir. Math. 8 (1967) 801–816 [2] Stability theorems for mappings with bounded excursion. Sibir. Math. 9 (1968) 667– 684 Rey, O. : [1] The role of the Green’s function in a non-linear elliptic equation involving the critical Sobolev exponent. J. Funct. Anal. 89, No. 1 (1990) 1–52 Riemann, B. : [1] Grundlagen f¨ ur eine allgemeine Theorie der Funktionen einer ver¨ anderlichen komplexen Gr¨ osse. Dissertation, G¨ ottingen (1851); Bernhard Riemann’s gesammelte mathematische Werke (ed. Weber, H.), Teubner, Leibzig (1876) 3–43 Rivi` ere, T. : [1] Flot des applications harmoniques en dimension deux. Preprint CMLA, No. 9223 (1992) Rockafellar, R.T. : [1] Convex analysis. Princeton University Press, Princeton (1970)

References

295

Rodemich, E. : [1] The Sobolev inequalities with best possible constants. In: Analysis Seminar, Cal. Inst. Tech. (1966) Rothe, E. : [1] Critical points and the gradient ﬁelds of scalars in Hilbert space. Acta Math. 85 (1951) 73–98 Rudin, W. : [1] Functional analysis. McGraw-Hill, New York (1973), 7th reprint, Tata McGraw-Hill, New Dehli (1982) Rupﬂin, M. : [1] An improved uniqueness result for the harmonic map ﬂow in two dimensions, preprint, ETH Z¨ urich (2007) Rybakowski, K.P. : [1] On the homotopy index for inﬁnite-dimensional semiﬂows. Trans. Am. Math. Soc. 269 (1982) 351–382 [2] The homotopy index on metric spaces with applications to partial diﬀerential equations. Freiburg (1985) Rybakowski, K.P.-Zehnder, E. : [1] On a Morse equation in Conley’s index theory for semiﬂows on metric spaces. Ergodic Theory Dyn. Syst. 5 (1985) 123–143 Sacks, P.-Uhlenbeck, K. : [1] On the existence of minimal immersions of 2-spheres. Ann. Math. 113 (1981) 1–24 Salamon, D. : [1] Connected simple systems and the Conley index of isolated invariant sets. Trans. Am. Math. Soc. 291 (1985) 1–41 Salvatore, A. : [1] Solutions of minimal period of a wave equation, via a generalization of Hofer’s theorem. Rend. Sem. Mat. Univ. Padova 81 (1989) 49–63 Schechter, M.-Tintarev, K. : [1] Spherical maxima in Hilbert space and semilinear elliptic eigenvalue problems. Diﬀ. Int. Eqs. 3 (1990) 889–899 Schoen, R.M. : [1] Conformal deformation of a Riemannian metric to constant scalar curvature. J. Diﬀ. Geom. 20 (1984) 479–495 Schoen, R.M.-Uhlenbeck, K. : [1] A regularity theory for harmonic maps. J. Diﬀ. Geom. 17 (1982) 307–335, and J. Diﬀ. Geom. 18 (1983) 329 [2] Boundary regularity and miscellaneous results on harmonic maps. J. Diﬀ. Geom. 18 (1983) 253–268 Schoen, R.M.-Yau, S.-T. : [1] Conformally ﬂat manifolds, Kleinian groups and scalar curvature. Invent. Math. 92 (1988) 47–71 Schoen, R.M.-Zhang, D. : [1] Prescribed scalar curvature on the n-sphere. Calc. Var. Partial Diﬀerential Equations 4 (1996), no. 1, 1–25 Schwartz, J.T. : [1] Generalizing the Lusternik-Schnirelmann theory of critical points. Commun. Pure Appl. Math. 17 (1964) 307–315 [2] Nonlinear functional analysis. Gordon & Breach, New York (1969) Schwarz, M. : [1] Morse homology. Progr. Math. 111, Birkh¨ auser, Basel-Boston-Berlin (1993) Schwetlick, H.-Struwe, M. : [1] Convergence of the Yamabe ﬂow for ”large” energies. J. Reine Angew. Math. 562 (2003), 59–100

296

References

Sedlacek, S. : [1] A direct method for minimizing the Yang-Mills functional over 4-manifolds. Commun. Math. Phys. 86 (1982) 515–527 Seifert, H. : [1] Periodische Bewegungen mechanischer Systeme. Math. Z. 51 (1948) 197–216 Semmes, S. : [1] A primer on Hardy spaces and some remarks on a theorem of Evans and M¨ uller. Commun. PDE 19 (1994) 277–319 Sere, E. : [1] Existence of inﬁnitely many homoclinic orbits in Hamiltonian systems. Math. Z. 209, No.1 (1992) 27–42 Serrin, J. : [1] On a fundamental theorem of the calculus of variations. Acta Math. 102 (1959) 1–22 [2] On the deﬁnition and properties of certain variational integrals. Trans. Am. Math. Soc. 101 (1961) 139–167 Shiﬀman, M. : [1] The Plateau problem for minimal surfaces of arbitrary topological structure. Am. J. Math. 61 (1939) 853–882 [2] The Plateau problem for non-relative minima. Ann. Math. 40 (1939) 834–854 [3] Unstable minimal surfaces with several boundaries. Ann. Math. 43 (1942) 197–222 Simader, C.G. : [1] On Dirichlet’s boundary value problem. Lect. Notes Math. 268, Springer, BerlinHeidelberg-New York (1972) Simon, L. : [1] Lectures on geometric measure theory. Centre Math. Analysis, Australian Nat. Univ. 3 (1983) [2] Asymptotics for a class of nonlinear evolution equations, with applications to geometric problems. Ann. Math. (2) 118 (1983), no. 3, 525–571 Smale, S. : [1] Generalized Poincar´ e’s conjecture in dimensions greater than four. Ann. Math. 74 (1961) 391–406 [2] Morse theory and a nonlinear generalization of the Dirichlet problem. Ann. Math. 80 (1964) 382–396 Solimini, S. : [1] On the existence of inﬁnitely many radial solutions for some elliptic problems. Revista Mat. Aplicadas 9 (1987) 75–86 Spanier, E.H. : [1] Algebraic topology. McGraw Hill, New York (1966) Steﬀen, K. : [1] On the existence of surfaces with prescribed mean curvature and boundary. Math. Z. 146 (1976) 113–135 [2] On the nonuniqueness of surfaces with prescribed constant mean curvature spanning a given contour. Arch. Rat. Mech. Anal. 94 (1986) 101–122 (received October 27, 1982) Struwe, M. : [1] Inﬁnitely many critical points for functionals which are not even and applications to superlinear boundary value problems. Manusc. math. 32 (1980) 335–364 [2] Multiple solutions of anticoercive boundary value problems for a class of ordinary diﬀerential equations of second order. J. Diﬀ. Eq. 37 (1980) 285–295 [3] Inﬁnitely many solutions of superlinear boundary value problems with rotational symmetry. Archiv d. Math. 36 (1981) 360–369 [4] Superlinear elliptic boundary value problems with rotational symmetry. Archiv d. Math. 39 (1982) 233–240 [5] A note on a result of Ambrosetti and Mancini. Ann. di Mat. 81 (1982) 107–115 [6] Multiple solutions of diﬀerential equations without the Palais-Smale condition. Math. Ann. 261 (1982) 399–412

References [7] [8] [9] [10] [11] [12] [13] [14] [15]

[16] [17] [18] [19]

[20] [21] [22] [23] [24] [25] [26] [27] [28]

297

Quasilinear elliptic eigenvalue problems. Comm. Math. Helv. 58 (1983) 509–527 A global compactness result for elliptic boundary value problems involving limiting nonlinearities. Math. Z. 187 (1984) 511–517 On a critical point theory for minimal surfaces spanning a wire in IRn . J. Reine Angew. Math. 349 (1984) 1–23 On the evolution of harmonic maps of Riemannian surfaces. Comm. Math. Helv. 60 (1985) 558–581 Large H-surfaces via the mountain-pass-lemma. Math. Ann. 270 (1985) 441–459 Nonuniqueness in the Plateau problem for surfaces of constant mean curvature. Arch. Rat. Mech. Anal. 93 (1986) 135–157, (received August 13, 1982) A Morse theory for annulus-type minimal surfaces. J. Reine Angew. Math. 368 (1986) 1–27 On the evolution of harmonic maps in higher dimensions. J. Diﬀ. Geom. 28 (1988) 485–502 Heat ﬂow methods for harmonic maps of surfaces and applications to free boundary problems. In: Partial Diﬀerential Equations (eds.: Cardoso, F.-de Figueiredo, D.G.I´ orio, R.-Lopes O.), Lect. Notes Math. 1324, Springer, Berlin (1988) The existence of surfaces of constant mean curvature with free boundaries. Acta Math. 160 (1988) 19–64 1,n

Critical points of embeddings of H0 into Orlicz spaces. Ann. Inst. H. Poincar´e, Analyse Non Lin´ eaire 5 (1988) 425–464 Plateau’s problem and the calculus of variations. Math. Notes 35, Princeton University Press, Princeton (1989) Multiple solutions to the Dirichlet problem for the equation of prescribed mean curvature, in: Analysis, et cetera (eds. Rabinowitz-Zehnder), Academic Press (1990) 639–666 Geometric evolution problems. Preprint 1992 and IAS/Park City Math. Ser. 2 (1996), 259–339 Existence of periodic solutions of Hamiltonian systems on almost every energy surface. Boletim Soc. Bras. Mat. 20 (1990) 49–58 Une estimation asymptotique pour le mod`ele Ginzburg-Landau. C. R. Acad. Sci. Paris 317 S´ er. 1 (1993) 677–680 On the asymptotic behavior of minimizers of the Ginzburg-Landau model in 2 dimensions. Diﬀ. Int. Eq. 7 (1994) 1613–1624 The evolution of harmonic maps. Proc. Intern. Congress Math. Kyoto 1990, SpringerVerlag, Tokyo (1991) 1197–1203 Positive solutions of critical semilinear elliptic equations on non-contractible planar domains. J. Eur. Math. Soc. (JEMS) 2 (2000), no. 4, 329–388 Curvature ﬂows on surfaces. Ann. Scuola Norm. Sup. Pisa (V) 1 (2002), 247–274 A ﬂow approach to Nirenberg’s problem. Duke Math. J. 128 (2005), no. 1, 19–64 Quantization for a fourth order equation with critical exponential growth, to appear in Math. Z. 256 (2007), No.2, 397–424

Struwe, M.-Tarantello, G. : [1] On multivortex solutions in Chern-Simons gauge theory. Bollettino U.M.I. 8 (1998) 109–121 Szulkin, A. : [1] Minimax principles for lower semicontinuous functions and applications to nonlinear boundary value problems. Ann. Inst. H. Poincar´e, Analyse Non Lin´ eaire, 3 (1986) 77–109 Talenti, G. : [1] Best constant in Sobolev inequality. Ann. Mat. Pura Appl. 110 (1976) 353–372 [2] Some inequalities of Sobolev-type on two-dimensional spheres. In: General inequalities 5 (ed.:Walter, W.), Intern. Ser. Numer. Math. 80, Birkh¨ auser (1987) 401–408

298

References

Tanaka, K. : [1] Inﬁnitely many periodic solutions for a superlinear forced wave equation. Nonlinear Analysis, T.M.A. 11 (1987) 85–104 Tartar, L.C. : [1] Compensated compactness and applications to partial diﬀerential equations. In: Nonlinear Analysis and Mechanics, Pitman Research Notes Math. 39, London (1979) 136–212 [2] The compensated compactness method applied to systems of conservation laws. In: Systems of nonlinear partial diﬀerential equations, Proc. NATO Advanced Study Inst. 111, Oxford 1982 (ed.: J.M. Ball), Reidel, Dordrecht (1983) 263–285 Thews, K. : [1] Non-trivial solutions of elliptic equations at resonance. Proc. Royal Soc. Edinburgh 85A (1980) 119–129 Tolksdorf, P. : [1] On the Dirichlet problem for quasilinear equations in domains with conical boundary points. Commun. P.D.E. 8 (1983) 773–817 [2] Regularity for a more general class of quasilinear elliptic equations. J. Diﬀ. Eq. 51 (1984) 126–150 Tomi, F.-Tromba, A.J. : [1] Extreme curves bound embedded minimal surfaces of the type of the disc. Math. Z. 158 (1978) 137–145 Tonelli, L. : [1] Fondamenti di calcolo delle variazioni. Zanichelli, Bologna (1921–23) Topping, P. : [1] Reverse bubbling and nonuniqueness in the harmonic map ﬂow. Int. Math. Res. Not. 2002, no. 10, 505–520 [2] Winding behaviour of ﬁnite-time singularities of the harmonic map heat ﬂow. Math. Z. 247 (2004) 279–302 [3] Repulsion and quantization in almost-harmonic maps, and asymptotics of the harmonic map ﬂow. Ann. Math. (2) 159 (2004) 465–534 Torchinsky, A. : [1] Real-variable methods in harmonic analysis. Academic Press, Orlando (1986) Trudinger, N.S. : [1] On embeddings into Orlicz spaces and some applications. J. Math. Mech. 17 (1967) 473–484 [2] Remarks concerning the conformal deformation of Riemannian structures on compact manifolds. Annali Sc. Norm. supp. Pisa 22 (1968) 265–274 Turner, R. : [1] Superlinear Sturm-Liouville problems. J. Diﬀ. Eq. 13 (1973) 157–171 Uhlenbeck, K. : [1] Regularity for a class of nonlinear elliptic systems. Acta Math. 138 (1977) 219–240 Vainberg, M.M. : [1] Variational methods for the study of non-linear operators. Holden Day, San Francisco (1964) Viterbo, C. : [1] A proof of the Weinstein conjecture in IR2n . Ann. Inst. H. Poincar´ e: Analyse non lin´ eaire 4 (1987) 337–356 [2] Indice de Morse des points critiques obtenus par minimax. Ann. Inst. H. Poincar´e, Analyse Non Lin´ eaire 5 (1988) 221–225 Walter, W. : [1] A theorem on elliptic diﬀerential inequalities with an application to gradient bounds. Math. Z. 200 (1989) 293–299 Wang, C. : [1] Bubble phenomena of certain Palais-Smale sequences from surfaces to general targets. Houston J. Math. 22 (1996) 559–590

References

299

Wang, G.-F. : [1] The Dirichlet problem for the equation of prescribed mean curvature. Ann. Inst. H. Poincar´ e, Analyse Non Lin´ eaire 9 (1992) 643–655 Wang, G.-F.-Wei, J. : [1] Steady state solutions of a reaction-diﬀusion system modeling chemotaxis. Math. Nachr. 233/234 (2002), 221–236 Wang, Z.Q. : [1] On a superlinear elliptic equation. Ann. Inst. H. Poincar´e: Analyse Non Lin´ eaire 8 (1991) 43–57 Weierstrass, K. : ¨ [1] Uber das sogenannte Dirichlet’sche Princip. Mathematische Werke II, Mayer & M¨ uller, Berlin (1895) 49–54 Weinstein, A. : [1] Normal modes for nonlinear Hamiltonian systems. Inv. Math. 20 (1973) 47–57 [2] Periodic orbits for convex Hamiltonian systems. Ann. Math. 108 (1978) 507–518 [3] On the hypotheses of Rabinowitz’ periodic orbit theorem. J. Diﬀ. Eq. 33 (1979) 353–358 Wente, H.C. : [1] An existence theorem for surfaces of constant mean curvature. J. Math. Anal. Appl. 26 (1969) 318–344 [2] A general existence theorem for surfaces of constant mean curvature. Math. Z. 120 (1971) 277–288 [3] The Dirichlet problem with a volume constraint. Manusc. math. 11 (1974) 141–157 [4] The diﬀerential equation Δx = 2Hxu ∧ xv with vanishing boundary values. Proc. Am. Math. Soc. 50 (1975) 59–77 [5] Large solutions to the volume constrained Plateau problem. Arch. Rat. Mech. Anal. 75 (1980) 59–77 Werner, H. : [1] Das Problem von Douglas f¨ ur Fl¨ achen konstanter mittlerer Kr¨ ummung. Math. Ann. 133 (1947) 303–319 Weth, T. : [1] Energy bounds for entire nodal solutions of autonomous superlinear equations. Calc. Var. Partial Diﬀerential Equations 27 (2006), no. 4, 421–437 Weyl, H. : ¨ [1] Uber die asymptotische Verteilung der Eigenwerte. Gesammelte Abhandlungen I (ed. Chandrasekharan, K.), Springer, Berlin-Heidelberg-New York (1968) 368–375 White, B. : [1] Inﬁma of energy functionals in homotopy classes of mappings. J. Diﬀ. Geom. 23 (1986) 127–142 Widman, K.O. : [1] H¨ older continuity of solutions of elliptic systems. Manusc. math. 5 (1971) 299–308 Willem, M. : [1] Remarks on the dual least action principle. Z. Anal. Anwendungen 1 (1982) 85–90 [2] Subharmonic oscillations of a semilinear wave equation. Nonlinear Analysis, T.M.A. 9 (1985) 503–514 Yamabe, H. : [1] On the deformation of Riemannian structures on compact manifolds. Osaka Math. J. 12 (1960) 21–37 Yang, C.T. : [1] On the theorems of Borsuk-Ulam. Kakutani-Yamabe-Yujobˇ o and Dysin. Part I: Ann. of Math. (2) 60 (1954) 262–282, part II: Ann. of Math. (2) 62 (1955) 271–280 Ye, R. : [1] Global existence and convergence of Yamabe ﬂow. J. Diﬀ. Geom. 39 (1994) 35–50

300

References

Zehnder, E. : [1] Periodische L¨ osungen von Hamiltonschen Systemen. Jber. d. Dt. Math.-Verein. 89 (1987) 33–59 [2] Remarks on periodic solutions on hypersurfaces. In: Periodic solutions of Hamiltonian systems and related topics (ed. Rabinowitz, P.H.), Reidel, Dordrecht (1987) 267–279 Zeidler, E. : [1] Vorlesungen u ¨ber nichtlineare Funktionalanalysis I–IV. Texte zur Math. 11, Teubner, Leipzig (1977) [2] Nonlinear functional analysis and its applications I–IV. Springer, New York-Berlin (1988-90)

Index

301

Index

Approximate solutions 33, 69 Area 6, 19, 220 Co-area formula 43 Barrier 273 (see also: Sub- and super-solution) Bifurcation 180 ﬀ. Cald´ eron-Zygmund inequality 268 f. Carath´ eodory function 275 Category 100 (see also: Index theory) Change in topology 169 Characteristic function 6 Coercive functional 4 operator 60 Compactness bounded compactness 2 concentration-compactness 37 f., 39 f., 44 f. compensated compactness 25 ﬀ. local compactness 41, 175, 226 global compactness in the critical case 184, 231 (see also: Palais-Smale condition) of a sequence of measures 39 Concentration-function (of measure) 40 Conformal conformal group of the disk 20 f. conformal invariance 20, 194, 232 conformal Laplace operator 197 conformality relation 20 ε-conformality theorem of Morrey 20 Convexity 9, 12 f., 21, 25, 54, 57, 61 f., 103, 105, 162 ﬀ. polyconvex 25 quasiconvex 54 Critical point (value) 1 at inﬁnity 169 in convex sets 164 of mountain pass type 143 of non-diﬀerentiable functional 152 saddle point 1, 50, 54, 73, 74, 76, 77, 87, 108, 151 Deformation lemma 81 ﬀ. for C 1 -functionals on Banach spaces 83 for C 1 -functionals on Finsler manifolds 87 for non-diﬀerentiable functionals 153 f., 158 on convex sets 165 Developing map 197 f.

Dichotomy (of a sequence of measures) 40 Dirichlet integral 20 Dual variational problem 62, 67 Eigenvalue for Dirichlet problem Courant-Fischer characterization 97 Rayleigh-Ritz characterization 14f. Weyl asymptotic formula 118 Elliptic equations 14 ﬀ., 16 ﬀ., 32 ﬀ., 98 ﬀ., 110 ﬀ., 116 ﬀ., 120 ﬀ., 128 f., 147 ﬀ., 150 ﬀ., 166 ﬀ., 170 ﬀ. degenerate elliptic equations 4 ﬀ., 183 on unbounded domains 36 ﬀ., 150 ﬀ. with critical growth 170 ﬀ. Energy energy functional 25, 71 f., 150, 231 energy inequality 199, 204, 238, 243 stored energy 25 energy surface 61 Epigraph 57 f. Equivariant 82, 84, 86, 94 Euler-Lagrange equations 1 Finsler manifolds 85 f. Fr´ echet-diﬀerential 274 Functional at inﬁnity 36 Genus 94 f. Geodesics 61, 89 closed geodesics on spheres 89 ﬀ. Gradient 84 gradient-ﬂow 84, 135 (see also: Pseudo-gradient) Group action 82, 84, 86, 94 Hamiltonian systems 60 ﬀ., 103 ﬀ., 124, 130 ﬀ., 137 ﬀ., 150 Hardy space 35, 224 Harmonic map 8, 71 ﬀ., 169, 231 ﬀ. evolution problem 238 ﬀ. Harmonic sphere 240 Index theory 94 ﬀ., 99 Benci-index 101 ﬀ. Krasnoselskii genus 94 ﬀ. Ljusternik-Schnirelman category 100 f. pseudo-index 101 Intersection lemma 113 Invariant under ﬂow 87 under group action (see: Equivariant) Isoperimetric inequality 43, 224 Legendre condition 13 Legendre-Fenchel transform 58 f., 63 f. Limiting problem 170, 184

302

Index

Linking 125 ﬀ. examples of linking sets 125 ﬀ., 134 Lower semi-continuity 2 ﬀ., 8 ﬀ., 25, 51, 58, 70 Maximum principle 271 ﬀ. Mean curvature equation 169, 220 ﬀ. Measure compactness of a sequence of measures 39 concentration function of measure 40 dichotomy of a sequence of measures 40 vanishing of a sequence of measures 39 Minimal surface 6 f., 19 ﬀ., 169 parametric minimal surface 19 f. minimal cones 7 minimal partitioning surfaces 6 f. Minimax principle 74, 87 ﬀ., 96 f. Courant-Fischer 97 Palais 87 Minimizer 1, 51, 70, 144, 166 Minimizing sequence 3, 53, 55 ﬀ., 70 Monotone operator 60 Monotonicity (of index) 99 Mountain pass lemma 74, 76, 108 ﬀ., 112 Palais-Smale condition 77 ﬀ. Condition (C) 78 Condition (P.-S.) 78 Cerami’s variant 80 for non-diﬀerentiable functionals 152 local 177 on convex sets 164 Palais-Smale sequence 54, 78 Penalty method 69 f. Periodic solutions of Hamiltonian systems 61 ﬀ., 103 ﬀ., 130 ﬀ., 137 ﬀ., 150 of semilinear wave equation 65 ﬀ., 124, 150 with prescribed minimal period 104 f. Perimeter (of a set) 6 Perron’s method 16 ﬀ. Plateau problem 19 ﬀ., 214 boundary condition 19

Pohoˇ zaev identity 155 f., 171 Poincar´ e inequality 266 f. Pseudo-gradient ﬂow 84 (see also: Deformation lemma) Pseudo-gradient vector ﬁeld 81, 86 for non-diﬀerentiable functionals 153 on convex sets 164 Pseudo-Laplace operator (p-Laplacian) 5, 183 Regular point (value) 1, 164 Regularity theory 16, 31, 57, 268 ﬀ. for minimal surfaces 24 for the constant mean curvature equation 223 in elasticity 31 partial regularity for evolution of harmonic maps 245 ﬀ., 248 f., 253 ﬀ. Rellich-Kondrakov theorem 265 Schauder estimates 268 Schwarz-symmetrization 42 Separation of spheres 169, 239 ﬀ. Sobolev embedding (inequality) 42 ﬀ., 170 ﬀ., 242, 265 ﬀ. density of smooth maps in Sobolev spaces 235 f. Sub-additivity (of index) 99 Sub-diﬀerential 58 Sub-solution 16 f. Super-solution 17 Supervariance (of index) 99 Support support hyperplane 58 support function 58 Symmetry 16, 36, 169 Symmetry group (see: Group action) Symplectic structure 60 Technique Fatou-lemma technique 34 hole-ﬁlling technique 56 Vanishing (of a sequence of measures) 39 Variational inequality 13, 166 Volume 220, 223 ﬀ. Wave equation 65 ﬀ., 124, 150 Yamabe problem 18, 194 ﬀ.